Boosting accuracy for efficient 2D stereo matching via attention-guided fusion and spatially adaptive modulation Authors: Zhong Wu, Lili He, Juan Li, Wenhuan Wu Published: 2026 ABSTRACT As a dense pixel-wise matching task, high-accuracy stereo matching generally incurs substantial computational overhead, leading to high deployment costs and inapplicability to many practical scenarios with limited computing resources or stringent low-latency demands. This has spurred a growing shift toward lightweight stereo matching models, yet such models remain far from satisfactory in accuracy, creating a notable accuracy gap. To address the pressing need for lightweight and high-accuracy stereo matching models, we propose LASM-Net, a lightweight stereo matching network with superior disparity accuracy. First, we discard the mainstream strategy of aggregating 4D cost volumes with 3D convolutions, which is computationally expensive, and instead adopt multi-scale 3D cost volumes and 2D convolutions, which feature low computational overhead. To mitigate the consequent accuracy degradation, spatially adaptive modulation matrices are learned to modulate the 3D cost volume across all three dimensions, thereby fully exploiting the disparity distribution cues and the spatial contextual information. Second, to curb excessive computational costs, existing lightweight models have to significantly reduce the number of convolutional layers, which greatly restricts the size of receptive fields and thus leads to a lack of contextual information. To address this dilemma, a multi-scale attention-guided fusion strategy is proposed, which effectively compensates for the lost contextual information with the help of low-resolution cost volumes. These key ideas, along with other tailored designs, are integrated to form our LASM-Net, which demonstrates superior performance in terms of both accuracy and inference speed, thereby narrowing the accuracy gap between lightweight and accuracy-oriented models significantly. Introduction Binocular stereo matching is one of the fundamental and critical tasks in various stereo vision applications and has been widely adopted in numerous vision-based systems, such as robotic navigation1,2, medical imaging3, 3D reconstruction4,5, and autonomous driving6,7. Given a pair of epipolar-rectified images, the goal of stereo matching is to find pixel-wise correspondences between the left (reference) and right (target) images; specifically, to generate a dense disparity map that characterizes these correspondences, from which the 3D scene structure can be readily recovered via triangulation8. For decades, recognized as the core but challenging task in stereo vision, stereo matching has attracted sustained and in-depth research efforts from many scholars9. Although traditional stereo matching methods have achieved great advancements, they typically rely on manually defined priors or assumptions (e.g., the local smoothness assumption). Such approaches face difficulties in simultaneously addressing the effects of various complex factors inherent in real-world scenarios, such as noise, illumination fluctuations, repeated textures, weak textures, and occlusions, ultimately leading to poor accuracy. In the past decade, deep learning-based stereo matching methods leverage the powerful learning capability of deep neural networks (DNNs) to address multiple challenges in stereo matching posed by various complex factors. Nowadays, their accuracy has significantly exceeded that of traditional methods, establishing them as the mainstream in the field. Unfortunately, as a pixel-wise dense matching task, the task of stereo matching inherently exhibits high computational complexity. Existing stereo matching networks with high accuracy typically require substantial computational overhead10–17, rendering them unsuitable for many practical scenarios with limited computing resources or strict low-latency requirements, such as edge-computing applications and real-time systems. Consequently, there has been a growing focus on the research and development of lightweight real-time stereo matching networks, which can roughly be divided into two categories according to their structural characteristics. One category of these models constructs and aggregates cost volumes at low resolutions, followed by super-resolution processing of either the resulting disparity maps18–23 or cost volumes24–26. Another category adopts a different approach: It constructs multi-scale cascaded cost volumes and focuses on reducing the number of disparity candidates at higher resolutions13,27–31. Despite the promising efficiency exhibited by these methodologies, their accuracy remains far from satisfactory, and a significant accuracy gap still persists when compared with accuracy-oriented models, as shown in Table 4 of Section "Comparisons on KITTI 2012 & KITTI 2015". In response to this long-standing dilemma, i.e., the inherent trade-off between efficiency and accuracy in lightweight stereo matching, we raise two critical questions that target the core: Is it even possible for speed-oriented lightweight models to achieve high accuracy? Further, what key components or mechanisms are missing from existing lightweight models that restrict their accuracy? As can be observed from existing works, most accuracy-oriented stereo matching networks construct 4D cost volumes via feature concatenation or group-wise correlation32, which generally encompass 4 dimensions regarding feature channels, disparity candidates, image height, and image width. Compared with 3D cost volumes, 4D cost volumes preserve more information from image feature maps during construction, enabling networks to perform exhaustive aggregation via 3D convolutions for accurate matching—at the cost of high memory consumption and prolonged runtime. In contrast, a few early studies33 adopted 3D cost volumes with 2D cost aggregation, where the cost volumes consist of three dimensions: disparity candidates, height, and width. These 3D cost volumes are typically built via feature correlation, resulting in relatively low computational overhead. However, compared with feature concatenation or group-wise correlation, feature correlation discards substantial information from image feature maps—making it challenging to achieve high matching accuracy with 3D cost volumes alone. Thus, the vast majority of contemporary speed-oriented lightweight models are compelled to adopt a compromise: constructing low-resolution 4D cost volumes while reducing the count of 3D convolutions. We argue that this compromise hinders lightweight models from achieving high accuracy for two key reasons. On the one hand, lightweight models must strictly constrain the resolution of 4D cost volumes and the count of 3D convolutional layers to ensure low computational complexity—this restriction impedes exhaustive aggregation and full capture of 3D spatial features. On the other hand, given the adoption of far fewer cascaded convolutional layers for feature extraction and cost aggregation, lightweight models exhibit relatively small receptive fields, resulting in a lack of broader contextual information. To mitigate these critical limitations and push the accuracy of lightweight models to a new benchmark, we draw inspiration from existing works18,23 and propose a lightweight framework that attains superior disparity accuracy with low computational overhead. First, we abandon the mainstream strategy of aggregating 4D cost volumes with 3D convolutions, which are computationally expensive, and instead embrace multi-scale 3D cost volumes and 2D convolutions, which feature low computational overhead. To compensate for the drawback of losing too much information in constructing 3D cost volumes, a spatially adaptive modulation (SAM) mechanism is adopted to modulate the 3D cost volume across all three dimensions (i.e., disparity, height, and width), thereby fully exploiting the disparity distribution cues and the spatial contextual information embedded in the volume. Second, as we cannot use more cascaded convolutional layers to obtain a wider range of contextual information due to computational overhead constraints, we depart from this approach. Instead, we construct a low-resolution 3D cost volume that can easily capture large-range contextual information while requiring only low computational overhead. Then, a multi-scale attention-guided fusion (MAGF) module is constructed to inject contextual information into the high-resolution 3D cost volume. These key ideas are integrated to finally form our framework for lightweight and accurate stereo matching, termed LASM-Net, which leverages merely 2D convolutions but demonstrates superior performance across accuracy, speed, and computational resource utilization, thereby significantly narrowing down the accuracy gap between speed-oriented and accuracy-oriented models. The excellent performance of LASM-Net suggests that the potential of the “3D cost volume + 2D convolution” paradigm has been somewhat underestimated in previous research efforts. The main contributions presented in this paper can be outlined as follows:A lightweight stereo matching framework termed LASM-Net is proposed, which is based on lightweight 3D cost volumes and 2D convolutions but achieves a strong balance between accuracy and inference speed. LASM-Net improves the accuracy-efficiency trade-off of real-time stereo matching models and helps narrow the accuracy gap between lightweight and accuracy-oriented models.A spatially adaptive modulation (SAM) mechanism is integrated into a lightweight encoder-decoder structure based on 2D convolutions for cost aggregation. By acting in synergy, they effectively compensate for the inherent limitation of excessive information loss during the construction of 3D cost volumes.An efficient multi-scale attention-guided fusion (MAGF) module is designed to inject contextual information from low-resolution cost volumes into high-resolution ones, which effectively remediates the deficiency that lightweight models struggle to obtain large-range contextual information. A lightweight stereo matching framework termed LASM-Net is proposed, which is based on lightweight 3D cost volumes and 2D convolutions but achieves a strong balance between accuracy and inference speed. LASM-Net improves the accuracy-efficiency trade-off of real-time stereo matching models and helps narrow the accuracy gap between lightweight and accuracy-oriented models. A lightweight stereo matching framework termed LASM-Net is proposed, which is based on lightweight 3D cost volumes and 2D convolutions but achieves a strong balance between accuracy and inference speed. LASM-Net improves the accuracy-efficiency trade-off of real-time stereo matching models and helps narrow the accuracy gap between lightweight and accuracy-oriented models. A spatially adaptive modulation (SAM) mechanism is integrated into a lightweight encoder-decoder structure based on 2D convolutions for cost aggregation. By acting in synergy, they effectively compensate for the inherent limitation of excessive information loss during the construction of 3D cost volumes. A spatially adaptive modulation (SAM) mechanism is integrated into a lightweight encoder-decoder structure based on 2D convolutions for cost aggregation. By acting in synergy, they effectively compensate for the inherent limitation of excessive information loss during the construction of 3D cost volumes. An efficient multi-scale attention-guided fusion (MAGF) module is designed to inject contextual information from low-resolution cost volumes into high-resolution ones, which effectively remediates the deficiency that lightweight models struggle to obtain large-range contextual information. An efficient multi-scale attention-guided fusion (MAGF) module is designed to inject contextual information from low-resolution cost volumes into high-resolution ones, which effectively remediates the deficiency that lightweight models struggle to obtain large-range contextual information. Related work As a fundamental but challenging task in computer vision, stereo matching has attracted sustained research attention for decades. Over the past decade, deep learning-based stereo matching methods have achieved remarkable progress and have gradually become the dominant paradigm. In this section, we review representative methods from an analytical perspective, focusing on four closely related aspects: cost-volume representation, cost-aggregation operator, contextual-information modeling, and the resulting accuracy-efficiency trade-off. Based on these aspects, existing methods can be broadly divided into accuracy-oriented models and speed-oriented models. This organization also helps clarify the technical position of the proposed LASM-Net relative to prior work. Accuracy-oriented stereo matching models Accuracy-oriented models focus on improving the accuracy of stereo matching through sophisticatedly designed modules and strategies, with relatively little concern for the model’s computational efficiency and inference latency. Based on the characteristics of different network structures, such models can be categorized into cost volume filtering-based models, iterative optimization-based models, and transformer-based models. Cost volume filtering-based models The vast majority of existing stereo matching methods based on deep learning fall into this category. These methods explicitly construct a cost volume using image features extracted by convolutional neural networks (CNNs); this cost volume typically encapsulates similarity information (or matching costs) between all potential matching pixel pairs across the entire set of pixels in the left and right images. Owing to the lack of robustness in feature similarity between individual pixel pairs, cost volume filtering (also known as cost aggregation) is generally required. GCNet34 is one of the earliest pioneering works that adopt 3D encoder-decoder architectures to regularize 4D concatenation volumes. In turn, PSMNet35 introduces a stacked hourglass 3D CNN with intermediate supervision embedded to regularize the 4D concatenation volume. GANet36 develops a local feature-guided aggregation layer and a semi-global aggregation layer to reduce the count of 3D convolutions while achieving further accuracy improvements. Meanwhile, GwcNet32 proposes the group-wise correlation volume, and ACVNet13,37 proposes an attention concatenation mechanism of cost volumes; both aim to enhance the representational capacity of cost volumes and improve the performance in ambiguous regions. To enhance cross-domain generalization performance, PCWNet38 incorporates warping cost volumes and a pyramid combination strategy. Nevertheless, the substantial computational and memory overhead of 4D cost volumes and 3D convolutions frequently restricts their application at higher resolutions. To address this memory inefficiency, a series of cascade architecture-based methods have been proposed10,13. Specifically, CFNet10 and CasStereo39 first construct a cost volume pyramid and then use a coarse-to-fine (CTF) strategy to gradually narrow down the search range of candidate disparities. Despite efforts to improve efficiency, this kind of method focuses on improving disparity accuracy and still demands high computational overhead. Iterative optimization-based models The models based on iterative optimization typically also construct cost volumes but incorporate an iterative optimization scheme into their frameworks to iteratively refine initial disparity estimations. Recently, such methods have achieved superior accuracy in stereo matching. As one of the pioneering works, RAFT40 proposes an iterative framework for optical flow estimation with high accuracy and robustness, leveraging convolutional gated recurrent units (GRUs) that perform lookups on 4D multi-scale correlation volumes to iteratively update the flow field. RAFT-Stereo41 extends this framework with multi-level convolutional GRUs and performs iterative updates of the initial disparity estimation using local cost values retrieved from a pyramid of 3D correlation volumes. CREStereo11 adopts a hierarchical structure with recurrent disparity refinement to perform coarse-to-fine disparity updates, whereas DLNR42 introduces a decoupled LSTM module to mitigate the loss of fine-grained details during the iterative process. IGEV-Stereo43 and IGEV++17 propose a combined geometry encoding volume (IGEV) and a multi-scale geometry-enhanced volume (MGEV) respectively, both of which aim at encoding geometry and context information and local matching details. Although iterative optimization-based methods have achieved high disparity accuracy with the aid of an iterative cost query and disparity update process, their iterative architecture inherently incurs increased computational overhead and prolonged inference latency. Transformer-based models Inspired by the great success of transformer-based models in various computer vision tasks, transformer-based models44–48 have emerged as an active research direction in stereo matching. A pioneering example is STTR44, which reformulates stereo correspondence as a sequence-to-sequence problem and employs transformers to replace the conventional cost volume built with fixed-range disparities. Leveraging this framework, CSTR47 boosts performance over STTR by integrating a context-enriched transformer design. Meanwhile, GMStereo46 exploits transformers to enable cross-view knowledge integration between left and right images via interactive mechanisms. Most recently, Croco-Stereo48 has achieved SOTA results by scaling up vision transformer-based cross-completion frameworks with extensive dataset utilization. However, due to the absence of explicitly built cost volumes, transformer-based approaches have to expand both their model scale and training dataset sizes to tap into the underlying semantic information of stereo matching, ultimately resulting in a considerably higher computational burden than the cost volume-based paradigm. Real-time stereo matching models Because many practical scenarios require stereo matching models to have low computational overhead, research on real-time stereo matching models has become an increasingly prominent trend in the field. Despite variations in the design details, existing real-time models can generally be categorized into two core paradigms based on the characteristics of volume construction and cost aggregation. The first paradigm typically constructs low-resolution 4-dimensional (4D) cost volumes—formed via feature concatenation or group-wise correlation—and aggregates them using 3D convolutions. StereoNet49 pioneered the end-to-end real-time framework for this paradigm, leveraging image-guided hierarchical disparity refinement to recover fine details in disparity maps. BGNet24, RTSMNet19, and LMNet21 adopt this core idea but introduce tailored upsampling strategies for improvement. Specifically, RTSMNet employs a feature-guided attention upsampling (FSAU) module to preserve details during the cascaded upsampling of disparity maps. LMNet performs multi-scale disparity refinement under image guidance. BGNet performs cost volume upsampling with the help of a learned bilateral grid, which upsamples low-resolution cost volumes via parameter-free slicing operations on the grid. CoEx18 contributes two key components: a feature-guided channel excitation mechanism for cost volumes (learning adaptive excitation weights to enhance geometric representations) and a top-k disparity regression strategy, which mitigates the inaccuracy of soft-argmin in ambiguous regions (e.g., textureless areas). Guided by image features, SAGU-Net26 generates adaptive attention weights for both cost volume and disparity map upsampling, effectively reducing detail loss during upsampling and enhancing the edge sharpness and detail integrity of the final disparity map. To further reduce computational overhead, unlike the StereoNet architecture, AnyNet27, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {RTS}^2$$\end{document}Net29, ADCPNet30, and SADSNet31 employ a coarse-to-fine (CTF) strategy. Specifically, they first build a low-resolution cost volume to derive an initial disparity map of low resolution, based on which a small set of new disparity candidates are generated for the construction of a compact high-resolution cost volume. Owing to the high spatial resolution but limited disparity range of the compact cost volume, computational overhead is significantly reduced while maintaining competitive accuracy. Fast-ACV13 designs a volume attention propagation submodule to recover high-resolution correlation volumes from low-resolution ones and selects candidates for high-likelihood disparity of the top-k to construct a compact concatenation volume. Since 4D cost volumes and 3D convolutions are inherently computationally expensive, the aforementioned methods primarily focus on pruning computational redundancy in their models via carefully designed frameworks and submodules. However, for real-time performance, such methods generally have to sacrifice their accuracy advantages significantly, yet they still struggle to achieve a better trade-off between accuracy and efficiency. The second paradigm embraces the computationally efficient 3D cost volumes and 2D convolutions. However, since the dimension of feature channels is eliminated when constructing 3D cost volumes, excessive information from image features is lost, resulting in relatively low accuracy. Consequently, such methods must employ diverse strategies to compensate for the information loss caused by 3D cost volumes. AANet50 constructs a pyramid of multi-scale 3D cost volumes from feature pyramids using cross-correlation, and alternately employs inter-scale cost aggregation and cross-scale cost fusion based on standard 2D convolutions and 2D deformable convolutions to mitigate information loss. FADNet25 adopts an encoder-decoder structure with dual residual blocks and multi-scale residual learning, combined with multi-round loss weight scheduling for coarse-to-fine training. P3SNet51 introduces pyramid pooling to capture multi-scale local/global features and hierarchical disparity aggregation to obtain multi-scale disparity estimations through a coarse-to-fine scheme. LightStereo23 leverages a pre-trained MobileNetV252 as its feature extraction backbone. For cost aggregation, it incorporates inverted residual blocks paired with a feature-directed channel excitation mechanism similar to CoEx, with a particular emphasis on exploiting channel-wise information embedded in 3D cost volumes. IINet53 proposes an intra-inter fusing framework that integrates intra-image and inter-image information, with a core focus on improving information flow in both feature and disparity domains. Although the elaborately designed modules and strategies in these 3D cost volume-based models have improved their disparity accuracy, a significant accuracy gap still exists between these models and accuracy-oriented approaches. Researchers in this field tend to argue that 3D cost volume-based methods, despite their low computational overhead, struggle to achieve high accuracy. However, we contend that existing methods have not yet fully exploited the potential of stereo matching paradigms based on 3D cost volumes and 2D convolutions. Instead of returning to computationally expensive 4D cost volumes and 3D convolutions, the proposed LASM-Net remains within the efficient “3D cost volume + 2D convolution” paradigm, but introduces two targeted mechanisms to address its major limitations. The proposed spatially adaptive modulation mechanism compensates for the information loss in 3D cost volumes by adaptively reweighting the cost volume along the disparity and spatial dimensions, while the proposed multi-scale attention-guided fusion module injects contextual cues from low-resolution cost volumes into high-resolution matching. Therefore, our LASM-Net aims to exploit the underexplored potential of this efficient paradigm and to achieve a stronger balance between accuracy and efficiency. Positioning of LASM-Net The above review suggests that existing stereo matching networks can be analyzed from four closely related perspectives: cost-volume representation, cost-aggregation operator, contextual-information modeling, and the resulting accuracy-efficiency trade-off. Table 1 summarizes the relationship between representative prior paradigms and our proposed LASM-Net. From this analytical comparison, LASM-Net is positioned as a lightweight stereo matching framework that specifically targets the key weaknesses of the “3D cost volume + 2D convolution” paradigm. Unlike accuracy-oriented models that rely on computationally expensive 4D cost volumes and 3D convolutions, LASM-Net maintains low computational overhead by using multi-scale 3D cost volumes and pure 2D operations. Unlike coarse-to-fine real-time methods that mainly reduce the disparity search range at high resolutions, LASM-Net aims to fully exploit the information contained in multi-scale cost volumes. Furthermore, compared with previous 3D cost-volume-based lightweight models, LASM-Net introduces two targeted mechanisms: spatially adaptive modulation, which compensates for information loss by adaptively weighting the cost volume along the disparity and spatial dimensions, and multi-scale attention-guided fusion, which injects wide-range contextual information from low-resolution cost volumes into high-resolution matching. Therefore, LASM-Net is not simply another efficient implementation of stereo matching, but a targeted attempt to narrow the accuracy gap between lightweight and accuracy-oriented models while preserving real-time efficiency.Table 1Analytical positioning of LASM-Net relative to representative stereo matching paradigms.ParadigmRepresentative methodsMain strength and limitationPosition of LASM-Net4D cost volume with 3D convolutionGCNet34, PSMNet35, GwcNet32, ACVNet13,37These methods preserve rich channel-wise matching information and achieve high accuracy, but their 4D volumes and 3D convolutions incur substantial memory and computational costs.LASM-Net avoids 4D cost volumes and 3D convolutions, thereby substantially reducing computational burden.Iterative optimization-based modelsRAFT-Stereo41, CREStereo11, IGEV-Stereo43, IGEV++17Iterative refinement improves disparity accuracy and robustness, but repeated cost lookup and update operations increase inference latency.LASM-Net adopts a feed-forward lightweight aggregation process rather than iterative refinement, making it more suitable for real-time deployment.Transformer-based modelsSTTR44, CSTR47, GMStereo46, Croco-Stereo48Transformer-based interaction can capture long-range cross-view dependencies, but usually requires larger model capacity and heavier computation.LASM-Net retains explicit cost-volume modeling and uses lightweight 2D operations to exploit context efficiently.Real-time models based on low-resolution or compact 4D volumesStereoNet49, BGNet24, CoEx18, AnyNet27, ADCPNet30, SADSNet31, Fast-ACV13These methods reduce redundancy through low-resolution aggregation, guided upsampling, or coarse-to-fine candidate pruning, but often sacrifice fine details or rely on coarse estimates.LASM-Net does not merely prune disparity candidates or refine low-resolution predictions; instead, it explicitly fuses low-resolution contextual cues with high-resolution cost volumes through MAGF.3D cost volume with 2D convolutionAANet50, FADNet25, P3SNet51, LightStereo23, IINet53This paradigm is computationally efficient, but correlation-based 3D volumes discard channel-wise feature information, which often limits accuracy.LASM-Net is positioned in this efficient paradigm, but further compensates for information loss via spatially adaptive modulation and enhances contextual modeling via multi-scale attention-guided fusion. Table 1Analytical positioning of LASM-Net relative to representative stereo matching paradigms.ParadigmRepresentative methodsMain strength and limitationPosition of LASM-Net4D cost volume with 3D convolutionGCNet34, PSMNet35, GwcNet32, ACVNet13,37These methods preserve rich channel-wise matching information and achieve high accuracy, but their 4D volumes and 3D convolutions incur substantial memory and computational costs.LASM-Net avoids 4D cost volumes and 3D convolutions, thereby substantially reducing computational burden.Iterative optimization-based modelsRAFT-Stereo41, CREStereo11, IGEV-Stereo43, IGEV++17Iterative refinement improves disparity accuracy and robustness, but repeated cost lookup and update operations increase inference latency.LASM-Net adopts a feed-forward lightweight aggregation process rather than iterative refinement, making it more suitable for real-time deployment.Transformer-based modelsSTTR44, CSTR47, GMStereo46, Croco-Stereo48Transformer-based interaction can capture long-range cross-view dependencies, but usually requires larger model capacity and heavier computation.LASM-Net retains explicit cost-volume modeling and uses lightweight 2D operations to exploit context efficiently.Real-time models based on low-resolution or compact 4D volumesStereoNet49, BGNet24, CoEx18, AnyNet27, ADCPNet30, SADSNet31, Fast-ACV13These methods reduce redundancy through low-resolution aggregation, guided upsampling, or coarse-to-fine candidate pruning, but often sacrifice fine details or rely on coarse estimates.LASM-Net does not merely prune disparity candidates or refine low-resolution predictions; instead, it explicitly fuses low-resolution contextual cues with high-resolution cost volumes through MAGF.3D cost volume with 2D convolutionAANet50, FADNet25, P3SNet51, LightStereo23, IINet53This paradigm is computationally efficient, but correlation-based 3D volumes discard channel-wise feature information, which often limits accuracy.LASM-Net is positioned in this efficient paradigm, but further compensates for information loss via spatially adaptive modulation and enhances contextual modeling via multi-scale attention-guided fusion. Analytical positioning of LASM-Net relative to representative stereo matching paradigms. Methodology In this section, the framework of the proposed model is first outlined, followed by a detailed description of its key components. Finally, the loss function employed for training is introduced. Overview of network structure Considering the need for stereo matching models to balance accuracy and efficiency, as well as the requirement to integrate local information and wide-range context for handling both detailed structures and low-texture regions, designing a lightweight stereo matching network involves intricate trade-offs and remains non-trivial. We draw inspiration from recent advances in stereo matching and other research fields. In accuracy-oriented stereo matching models10, multi-scale features and multi-scale cost volumes have proven critical for integrating local information and wide-range context, and are thus widely adopted in such models. However, in existing lightweight models with multi-scale structures30,31, low-resolution cost volumes are typically used to generate initial disparity estimates for building high-resolution cost volumes, and thus substantial information contained in low-resolution cost volumes is lost, thereby preventing the full exploitation of the multi-scale framework’s potential. Additionally, in the field of superpixel segmentation55 and image super-resolution56, context-aware upsampling is critical and fully exploited. Thus, we perform upsampling on the predicted disparity map to its full resolution through a lightweight context-aware upsampling scheme. With these insights, we design a Lightweight and Accurate Stereo Matching Network (LASM-Net). As shown in Fig. 1, the proposed LASM-Net comprises three key components: a pretrained multi-scale feature extractor (MFE) for image feature extraction, an encoder-decoder with spatially adaptive modulation (ED-SAM) for cost volume aggregation, and a multi-scale attention-guided fusion (MAGF) mechanism for multi-scale cost volume fusion. First, the multi-scale feature extractor (MFE) extracts feature maps at different resolutions from the input stereo images. Second, two 3D cost volumes are constructed through correlation at different scales. Third, the low-resolution cost volume is regularized by ED-SAM, where the encoder-decoder structure aggregates multi-scale contextual information and the SAM module injects feature-guided modulation into the cost volume to compensate for the information loss caused by correlation-based 3D cost-volume construction. Finally, the regularized low-resolution cost volume and the high-resolution cost volume are adaptively fused by MAGF, where a learned attention matrix balances wide-range contextual information and local structural details. We point out that the architecture of LASM-Net actually follows three design principles rather than being a collection of independent heuristic modules. First, the model should avoid the expensive “4D cost volume + 3D convolution” paradigm, since both memory consumption and computational complexity grow rapidly with the feature-channel, disparity, height, and width dimensions. Second, the model should compensate for the information loss caused by correlation-based 3D cost volumes, where two high-dimensional feature vectors are compressed into a scalar matching response. Third, the model should preserve both local structural details and wide-range contextual information under a strict computational budget. Generally, our design aims to fully unleash the potential of the multi-scale framework based on 3D cost volumes and 2D convolutions while demanding low computational overhead.Fig. 1The proposed lightweight and accurate stereo matching network (LASM-Net) comprises three key components: the multi-scale feature extractor (MFE), the encoder-decoder with spatially adaptive modulation (ED-SAM) for cost aggregation, and the multi-scale attention-guided fusion (MAGF) mechanism for multi-scale cost volume fusion. Our design aims to fully unleash the potential of the multi-scale framework while maintaining low computational overhead. The proposed lightweight and accurate stereo matching network (LASM-Net) comprises three key components: the multi-scale feature extractor (MFE), the encoder-decoder with spatially adaptive modulation (ED-SAM) for cost aggregation, and the multi-scale attention-guided fusion (MAGF) mechanism for multi-scale cost volume fusion. Our design aims to fully unleash the potential of the multi-scale framework while maintaining low computational overhead. Multi-scale feature extractor (MFE) Both high-level semantic context and low-level perceptual details are crucial for stereo matching, where the former facilitates matching in low-texture regions and the latter preserves delicate structures. Therefore, following previous research18,23,50, we design the MFE module to capture high-level context with large receptive fields as well as low-level local features. The structure of MFE is illustrated in the left part of Fig. 1. Specifically, we first adopt MobileNetV252—a classic mobile architecture originally designed for object detection and semantic segmentation—to rapidly downsample spatial resolutions and extract feature maps with multi-scale semantic contexts. Next, a decoder structure integrated with a feature fusion module (FFM) is employed to upsample spatial resolutions and fuse feature maps at different depths. The structure of FFM is illustrated in Fig. 2. Finally, feature maps of four scales are obtained, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L/R,i} \in \mathbb {R}^{C_i \times H_i \times W_i}$$\end{document} for i = 1,2,3,4. Here, the subscripts L and R are used to distinguish feature maps extracted from the left and right images, respectively, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_i$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_i$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_i$$\end{document} represent the number of channels, height, and width of the corresponding feature maps. In this paper, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_i$$\end{document} are set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H/2^{i+1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W/2^{i+1}$$\end{document}, respectively, where H × W denotes the resolution of the input stereo pair.Fig. 2Structure of the feature fusion module (FFM). Conv and DeConv denote the standard 2D convolution and 2D deconvolution, respectively. i, o, k, and s indicate input and output channels, kernel size, and stride, respectively. Structure of the feature fusion module (FFM). Conv and DeConv denote the standard 2D convolution and 2D deconvolution, respectively. i, o, k, and s indicate input and output channels, kernel size, and stride, respectively. With feature maps \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L/R,1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L/R,2}$$\end{document}, we build two cost volumes of two different scales via cross-correlation, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_2$$\end{document}, respectively. Take \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_1$$\end{document} as an example, mathematically,1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{C}_1(d, y, x)=\frac{1}{C_1}\left\langle \boldsymbol{F}_{L,1}(y, x), \boldsymbol{F}_{R,1}(y, x - d)\right\rangle \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\langle \cdot , \cdot \rangle$$\end{document} denotes the inner product, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L,1}(y, x)$$\end{document} represents the feature vector of pixel (x, y) in the left image, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d \in [0,1,2,...,D_1-1]$$\end{document} is a candidate disparity, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_1$$\end{document} is the count of candidate disparities, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_1 \in \mathbb {R}^{D_1 \times H_1 \times W_1}$$\end{document}. With an equation similar to Eq. 1, the cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_2 \in \mathbb {R}^{D_2 \times H_2 \times W_2}$$\end{document} can be obtained. Note that since both \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_2$$\end{document} are 3D cost volumes, they require relatively small memory footprints and can be aggregated with efficient 2D convolutions. However, we argue that because the inner product takes two vectors as input but outputs a scalar, too much information from the feature maps (i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L/R,1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L/R,2}$$\end{document}) is lost during the construction of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_2$$\end{document}. Thus, it is difficult to achieve satisfactory results merely using these 3D cost volumes. To address this issue, inspired by several prior works18, we incorporate feature maps of the left image into the cost aggregation process to compensate for the lost feature information. Cost aggregation As shown in Fig. 1, the cost aggregation pipeline of our proposed model is streamlined and straightforward. Specifically, under the guidance of image features, the low-resolution 3D cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_2$$\end{document} is regularized by the proposed ED-SAM module which is actually a 2D encoder-decoder (ED) structure equipped with spatially adaptive modulation (SAM), where the ED structure is designed to effectively aggregate multi-scale context and the SAM mechanism is designed to compensate for the lost feature information. Then, an efficient multi-scale attention-guided fusion module (MAGF) is employed to fuse the high-resolution cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_1$$\end{document} and the regularized low-resolution cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{C}}_2$$\end{document}. Encoder-decoder with spatially adaptive modulation (ED-SAM) The structure of ED-SAM is illustrated in Fig. 3, which is built on inverted residual blocks52 and channel excitation18,23. Guided by the multi-scale features of the left image, ED-SAM performs multi-scale cost aggregation via an encoding-decoding process.Fig. 3Structure of the encoder-decoder (ED) with spatially adaptive modulation (SAM) for cost aggregation. IRB-T1 and IRB-T2 represent the inverted residual blocks (IRBs) of type 1 and type 2, respectively. SAM and Deconv represent the spatially adaptive modulation module and 2D transposed convolution, respectively. a, b and c indicate the number of IRBs at each stage of the encoder. Structure of the encoder-decoder (ED) with spatially adaptive modulation (SAM) for cost aggregation. IRB-T1 and IRB-T2 represent the inverted residual blocks (IRBs) of type 1 and type 2, respectively. SAM and Deconv represent the spatially adaptive modulation module and 2D transposed convolution, respectively. a, b and c indicate the number of IRBs at each stage of the encoder. Specifically, as illustrated in Fig. 3, the encoding process comprises three stages operating at distinct scales, where each stage is composed of several cascaded inverted residual blocks (IRBs) and a spatially adaptive modulation module (SAM). As shown in Fig. 4, the IRB comprises two variants with slightly different structures: IRB-T1 incorporates a residual connection and outputs cost volumes with the same dimensions as its input; in contrast, IRB-T2 abandons the residual connection and outputs cost volumes with a doubled channel count and halved spatial resolution (i.e., halved height and width). The core design principle of IRBs is to first expand the channel dimension (i.e., the disparity dimension of cost volumes), subsequently perform depthwise convolutions, and finally shrink the channel dimension back—a design that effectively enhances feature representation while maintaining low computational overhead.Fig. 4Detailed structures of the inverted residual blocks of type 1 (IRB-T1) and type 2 (IRB-T2). Conv and BN denote the standard 2D convolution and batch normalization, respectively, and ReLU6 is the activation function. The parameters i, o, k, s and g indicate the input and output channel counts, kernel size, stride and group count of standard 2D convolutions, respectively. Detailed structures of the inverted residual blocks of type 1 (IRB-T1) and type 2 (IRB-T2). Conv and BN denote the standard 2D convolution and batch normalization, respectively, and ReLU6 is the activation function. The parameters i, o, k, s and g indicate the input and output channel counts, kernel size, stride and group count of standard 2D convolutions, respectively. To compensate for the information loss in building 3D cost volumes, multi-scale image features are injected into the cost volume via the spatially adaptive modulation module (SAM) at the end of each encoding stage. The details of SAM are shown in Fig. 5. Concretely, an attention matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{A}$$\end{document} is first estimated from the input feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L,\text {in}}$$\end{document} through two standard point-wise convolutions and several depthwise asymmetric convolutions with large kernel sizes, which can effectively aggregate multi-scale context at very low computational overhead. Then, the attention matrix is applied to modulate the input cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_{in}$$\end{document}. Mathematically, the modulated cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_{out}$$\end{document} is calculated as2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\boldsymbol{X}_1 = Conv_{1\times 1} (\boldsymbol{F}_{L,in}) \end{aligned}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\boldsymbol{X}_2 = \frac{1}{4} \big [ {DW}_{21\times 1}( {DW}_{1\times 21}(\boldsymbol{X}_1) ) + {DW}_{11\times 1}( {DW}_{1\times 11}(\boldsymbol{X}_1) ) + {DW}_{7\times 1}( {DW}_{1\times 7}(\boldsymbol{X}_1) ) +\boldsymbol{X}_1 \big ] \end{aligned}$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\boldsymbol{A} = Softmax \left( Conv_{1\times 1} (\boldsymbol{X}_2) \right) \end{aligned}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\boldsymbol{C}_{out} = \boldsymbol{A} \odot \boldsymbol{C}_{in} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Conv_{1\times 1}(\cdot )$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$DW_{1\times 21}(\cdot )$$\end{document}, Softmax(· ) and ⊙ represent 1× 1 convolution (i.e., point-wise convolution), depthwise convolution with asymmetric 1× 21 kernel, softmax operation, and element-wise multiplication, respectively. After the encoding process, the decoding module of ED-SAM progressively upsamples the spatial resolution of the encoded cost volume using transposed convolutions, which are coupled with skip connections. These skip connections serve as efficient propagation pathways for shallow features in the encoding module, as illustrated in Fig. 3. We point out that, although constructing and aggregating 3D cost volumes requires significantly lower computational overhead than 4D cost volumes, the correlation operation used to build 3D cost volumes compresses channel-wise feature interactions into a scalar similarity value, leading to considerable information loss. Note that a 3D cost volume is in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {R}^{D \times H \times W}$$\end{document} and a 4D volume is in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {R}^{C \times D \times H \times W}$$\end{document}. This explains why 3D volume-based models typically enjoy faster inference speed yet deliver relatively lower accuracy. These observations explain the effectiveness of our proposed ED-SAM: it introduces feature-guided modulation into the cost volume and enables adaptive weighting along the disparity and spatial dimensions, thereby compensating for the information lost in 3D cost volumes to some extent.Fig. 5Structure of the spatially adaptive modulation module (SAM). “Conv” denotes the standard 2D convolution, where i, o, k and g indicate the input and output channel counts, kernel size and group count, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[D_{in},H_{in},W_{in}]$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[C_{in},H_{in},W_{in}]$$\end{document} indicate the dimensions of the input cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_{in}$$\end{document} and feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L,in}$$\end{document}, respectively. “Mul” denotes element-wise multiplication. Structure of the spatially adaptive modulation module (SAM). “Conv” denotes the standard 2D convolution, where i, o, k and g indicate the input and output channel counts, kernel size and group count, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[D_{in},H_{in},W_{in}]$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[C_{in},H_{in},W_{in}]$$\end{document} indicate the dimensions of the input cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_{in}$$\end{document} and feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L,in}$$\end{document}, respectively. “Mul” denotes element-wise multiplication. Multi-scale attention-guided fusion module (MAGF) Integrating the advantages of low-resolution cost volumes (encompassing wide-range context) and high-resolution cost volumes (capturing local details) is critical for stereo matching models to simultaneously ensure disparity accuracy in regions with intricate structures and those with low texture. However, it is non-trivial to efficiently fuse cost volumes with distinct resolutions—each encodes scale-specific information, and thus one may undermine the other’s valuable information. To address this issue efficiently, a lightweight and effective multi-scale attention-guided fusion module (MAGF) is constructed to endow the fusion process with self-adaptive capability. As shown in Fig. 6, a spatially adaptive attention matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{B} \in \mathbb {R}^{D_1 \times H_1 \times W_1}$$\end{document} is learned from the input feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F} \in \mathbb {R}^{C_1 \times H_1 \times W_1}$$\end{document} by 1 × 1 convolutions and depthwise asymmetric convolutions with large kernel sizes. The low-resolution cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{C}}_2 \in \mathbb {R}^{D_2 \times H_2 \times W_2}$$\end{document} and high-resolution cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_1 \in \mathbb {R}^{D_1 \times H_1 \times W_1}$$\end{document} are then fused under the guidance of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{B}$$\end{document}. Mathematically, the resulting cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\tilde{C}}_1$$\end{document} is calculated as6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\boldsymbol{Y}_1 = Conv_{1 \times 1} (\boldsymbol{F}_{L,1}) \end{aligned}$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\boldsymbol{Y}_2 = \frac{1}{4} \big [ {DW}_{21 \times 1}( {DW}_{1 \times 21}(\boldsymbol{Y}_1) ) + {DW}_{11 \times 1}( {DW}_{1 \times 11}(\boldsymbol{Y}_1) ) + {DW}_{7 \times 1}( {DW}_{1 \times 7}(\boldsymbol{Y}_1) ) + \boldsymbol{Y}_1 \big ] \end{aligned}$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\boldsymbol{B} = Softmax \left( Conv_{1 \times 1} (\boldsymbol{Y}_2) \right) \end{aligned}$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\boldsymbol{\tilde{C}}_1 = \boldsymbol{B} \odot \boldsymbol{C}_1 + (1-\boldsymbol{B}) \odot BI(\boldsymbol{\hat{C}}_2) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Conv_{1 \times 1}(\cdot )$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$DW_{1 \times 21}(\cdot )$$\end{document}, Softmax(· ) and ⊙ denote 1 × 1 convolution (i.e., point-wise convolution), depthwise convolution with an asymmetric 1 × 21 kernel, the softmax operation, and element-wise multiplication, respectively. BI represents bilinear interpolation.Fig. 6Details of our proposed multi-scale attention-guided fusion module (MAGF). Conv denotes the standard 2D convolution, where i, o, k and g indicate the input and output channel counts, kernel size and group count, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[D_1,H_1,W_1]$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[D_2,H_2,W_2]$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[C_1,H_1,W_1]$$\end{document} indicate dimensions of the input high-resolution cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_1$$\end{document}, the input low-resolution cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\boldsymbol{C}_2}$$\end{document}, and feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L,1}$$\end{document}, respectively. BI and Mul represent bilinear interpolation and element-wise multiplication, respectively. Details of our proposed multi-scale attention-guided fusion module (MAGF). Conv denotes the standard 2D convolution, where i, o, k and g indicate the input and output channel counts, kernel size and group count, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[D_1,H_1,W_1]$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[D_2,H_2,W_2]$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[C_1,H_1,W_1]$$\end{document} indicate dimensions of the input high-resolution cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{C}_1$$\end{document}, the input low-resolution cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\boldsymbol{C}_2}$$\end{document}, and feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{F}_{L,1}$$\end{document}, respectively. BI and Mul represent bilinear interpolation and element-wise multiplication, respectively. Since low-resolution cost volumes provide broader context while high-resolution cost volumes preserve finer structures, MAGF performs adaptive fusion of different cost volumes rather than fixed fusion, allowing the network to balance context and local detail according to spatially varying image content. Note that the attention matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{B} \in \mathbb {R}^{D_1 \times H_1 \times W_1}$$\end{document} is learned from high-resolution feature maps via multi-scale asymmetric large-kernel convolutions, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{B}$$\end{document} has the same dimensions as the high-resolution cost volume. The MAGF enables adaptive fusion along all three dimensions (disparity, height, width). This effectively mitigates the adverse impacts of fusing scale-specific information across varying resolutions, thereby mitigating the degradation of valuable information in the fused cost volume. Theoretical analysis of computational complexity We further analyze the computational complexity of the whole ED-SAM module according to its implementation. Specifically, ED-SAM is realized by a three-stage encoder-decoder architecture. The encoder contains a, b, and c inverted residual blocks (IRBs) at the three stages, respectively, where the second and third stages each start with a stride-2 IRB for downsampling. The decoder contains two transposed convolutions and two redirection branches. Moreover, each encoder stage is followed by one SAM module. Let the input 3D cost volume of ED-SAM be in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {R}^{D_1\times H_1\times W_1}$$\end{document}, and let r denote the expansion ratio in the inverted residual block. The three encoder stages then process feature volumes of sizes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_1\times H_1\times W_1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2D_1\times H_1/2\times W_1/2$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4D_1 \times H_1/4 \times W_1/4$$\end{document}, respectively. For an IRB with input channel number \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{\textrm{in}}$$\end{document}, output channel number \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{\textrm{out}}$$\end{document}, spatial size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_s\times W_s$$\end{document}, and expansion ratio r, the dominant computations come from: 1) a 1× 1 expansion convolution, 2) a 3× 3 depthwise convolution, and 3) a 1× 1 projection convolution. For an IRB-T1 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{\textrm{out}}=C_{\textrm{in}}$$\end{document}, its complexity is10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{O}_{\mathrm {IRB\text{- }T1}}(C_{\textrm{in}},H_s,W_s) = \boldsymbol{O}\!\left( (2rC_{\textrm{in}}^2 + 9rC_{\textrm{in}})H_sW_s\right) \approx \boldsymbol{O}(rC_{\textrm{in}}^2H_sW_s). \end{aligned}$$\end{document}For an IRB-T2 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{\textrm{out}}=2C_{\textrm{in}}$$\end{document}, its complexity is11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{O}_{\mathrm {IRB\text{- }T2}}(C_{\textrm{in}},H_s,W_s) = \boldsymbol{O}\!\left( \frac{3}{2}rC_{\textrm{in}}^2H_sW_s + \frac{9}{4}rC_{\textrm{in}}H_sW_s\right) \approx \boldsymbol{O}(rC_{\textrm{in}}^2H_sW_s). \end{aligned}$$\end{document}Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(2D_1)^2 \cdot H_1 W_1/4=D_1^2H_1W_1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(4D_1)^2 \cdot H_1 W_1 /16=D_1^2 H_1 W_1$$\end{document}, the dominant complexity of IRBs remains at the same order across the three stages after resolution scaling. Therefore, the total dominant complexity of all encoder IRBs can be written as12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{O}_{\mathrm {enc\mathrm{- }IRB}} = \boldsymbol{O}\!\left( \big [2(a+b+c)-1\big ]\,rD_1^2H_1W_1\right) , \end{aligned}$$\end{document}where the coefficient comes from (a+b+c-2) stride-1 IRBs and two stride-2 IRBs. In addition, the two redirection branches are also implemented with IRBs, and their total dominant complexity is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{O}_{\textrm{redir}}=\boldsymbol{O}(4rD_1^2H_1W_1)$$\end{document}. For the decoder, the two transposed convolutions have complexities13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{O}\!\left( 9\cdot 4D_1\cdot 2D_1\cdot H_1W_1/4 \right) =\boldsymbol{O}(18D_1^2H_1W_1), \quad \text {and} \quad \boldsymbol{O}\!\left( 9\cdot 2D_1\cdot D_1 \cdot H_1W_1\right) =\boldsymbol{O}(18D_1^2H_1W_1), \end{aligned}$$\end{document}respectively. Thus, the total decoder complexity is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{O}_{\textrm{dec}}=\boldsymbol{O}(36D_1^2H_1W_1)$$\end{document}. Now consider the three SAM modules. Let the guiding feature channel numbers at the three stages be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_2$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_3$$\end{document}, respectively. Then the total SAM complexity is14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{O}_{\textrm{3SAM}}&= \boldsymbol{O}\!\left( (C_1D_1 + D_1^2)H_1W_1\right) + \boldsymbol{O}\!\left( \left( C_2\cdot 2D_1 + (2D_1)^2\right) \frac{H_1W_1}{4}\right) + \boldsymbol{O}\!\left( \left( C_3\cdot 4D_1 + (4D_1)^2\right) \frac{H_1W_1}{16}\right) \end{aligned}$$\end{document}15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&= \boldsymbol{O}\!\left( \left[ 3D_1^2 + \left( C_1 + \frac{1}{2}C_2 + \frac{1}{4}C_3\right) D_1\right] H_1W_1\right) . \end{aligned}$$\end{document}Combining all components, the total complexity of the whole ED-SAM is16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{O}_{\mathrm {ED\mathrm{- }SAM}} = \boldsymbol{O}_{\mathrm {enc\mathrm{- }IRB}} + \boldsymbol{O}_{\textrm{redir}} + \boldsymbol{O}_{\textrm{dec}} + \boldsymbol{O}_{\textrm{3SAM}}, \end{aligned}$$\end{document}namely,17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{O}_{\mathrm {ED\mathrm{- }SAM}} = \boldsymbol{O}\!\left( \left( [2(a+b+c)+3]r + 39 \right) D_1^2H_1W_1 + \left( C_1 + \frac{1}{2}C_2 + \frac{1}{4}C_3\right) D_1H_1W_1 \right) . \end{aligned}$$\end{document}Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_1$$\end{document} is typically not small in stereo matching (a typical value is 192/4=48), the quadratic term dominates in practice, and the overall complexity can be simplified as18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{O}_{\mathrm {ED\text{- }SAM}} \approx \boldsymbol{O}\!\left( \left( [2(a+b+c)+3]r + 39\right) D_1^2H_1W_1\right) . \end{aligned}$$\end{document}For comparison, if one instead regularizes a conventional 4D cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {R}^{C\times D_1\times H_1\times W_1}$$\end{document} using L layers of 3D convolutions with kernel size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K\times K\times K$$\end{document}, the corresponding complexity is19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{O}_{\mathrm {4D+3DConv}} = \boldsymbol{O}\!\left( LC^2K^3D_1H_1W_1\right) . \end{aligned}$$\end{document}Under the typical settings a=2, b=4, c=8, r=4, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_1=48$$\end{document}, K=3, C=16, and L=5, the complexity ratio between the conventional “4D cost volume + 3D convolution” scheme and ED-SAM is20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \frac{O_{\mathrm {4D+3DConv}}}{O_{\mathrm {ED\text{- }SAM}}} = \frac{144L}{163} \approx 4.42. \end{aligned}$$\end{document}Therefore, the conventional paradigm is about 4.42× more computationally expensive than ED-SAM under this setting. The conventional paradigm also becomes more expensive when more 3D convolution layers are used, and its computational cost increases linearly with L. This further demonstrates the efficiency advantage of ED-SAM for practical multi-layer cost aggregation. The above theoretical analysis shows that ED-SAM avoids the expensive 3D filtering over the disparity-height-width space of a 4D cost volume. Instead, it performs cost aggregation on lightweight 3D cost volumes using only 2D operators and attention-guided modulation. This is precisely why ED-SAM can strengthen cost regularization while still maintaining substantially lower computational complexity than the mainstream “4D cost volume + 3D convolution” paradigm. Disparity regression After cost aggregation, a disparity map can be derived from each of the resulting cost volumes. For the low-resolution cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{C}}_2$$\end{document}, the predicted disparity map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{D}}_2$$\end{document} is calculated using the soft argmin operation34. Specifically, the predicted disparity for each pixel is computed as the probability-weighted sum of all disparity candidates. Concretely,21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{\hat{D}}_2(y, x) = \sum _{d = 0}^{D_2 - 1}{d\times softmax \left( \boldsymbol{\hat{C}}_2(d, y, x) \right) } \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$softmax \left( \cdot \right)$$\end{document} denotes converting the matching cost into a probability through the softmax operation along the disparity dimension, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{D}}_2 \in \mathbb {R}^{H_2 \times W_2}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{C}}_2 \in \mathbb {R}^{D_2 \times H_2 \times W_2}$$\end{document}. For the high-resolution cost volume (i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{C}}_1$$\end{document}), the Top-k disparity regression strategy is adopted, where each disparity estimate is calculated as the probability-weighted sum of k disparity candidates with the highest matching scores (i.e., the k most likely candidates). This strategy is helpful when the matching cost distribution is relatively uniform or exhibits multiple peaks18. Mathematically,22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{\hat{D}}_1(y, x) = \sum _{d_i \in \boldsymbol{S}_{\text {top-}k} (y, x) } {d_i \times softmax \left( \boldsymbol{\hat{C}}_1(d_i, y, x) \right) } \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{S}_{\text {top-}k} (y, x)$$\end{document} represents the set of k disparity candidates that have the highest matching scores (i.e., the lowest matching costs), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{D}}_1 \in \mathbb {R}^{H_1 \times W_1}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{C}}_1 \in \mathbb {R}^{D_1 \times H_1 \times W_1}$$\end{document}. Finally, following18,55, the full-resolution disparity map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{D}}_{full} \in \mathbb {R}^{H \times W}$$\end{document} is obtained from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{D}}_1$$\end{document} using a context-aware upsampling strategy. Specifically, the disparity for each pixel in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{D}}_{full}$$\end{document} is computed as a weighted aggregation over its \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\!\times \!3$$\end{document} superpixel neighborhood in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\hat{D}}_1$$\end{document}, where the weights are learned from image feature maps through a series of 2D convolutions. Training loss To effectively train the proposed LASM-Net, we supervise the full-, high-, and low-resolution disparity maps simultaneously. Mathematically, the loss function is formulated as23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&L = \sum _{m = 1}^{3} \lambda _m T_m, \end{aligned}$$\end{document}24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&T_m = \frac{1}{P_m} \sum _{(x, y)} \text {smooth}_{L_1} \big ( \boldsymbol{D}_{gt}(y, x) - \boldsymbol{D}_m(y, x) \big ) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{D}_m \in \{ \boldsymbol{\hat{D}}_1,\boldsymbol{ \hat{D}}_2, \boldsymbol{\hat{D}}_{full}\}$$\end{document} denotes the m-th predicted disparity map, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{D}_{gt}$$\end{document} represents the ground-truth disparity map, (x, y) represent pixel coordinates, the scalar \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _m$$\end{document} is used to balance the importance of each loss term, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$smooth_{L_1}(\cdot )$$\end{document} is the smooth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_1$$\end{document} function, and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_m$$\end{document} denotes the number of valid pixels. Experiments Datasets and evaluation metrics Scene Flow33 is a large-scale synthetic stereo benchmark, which comprises 35,454 training and 4,370 testing stereo image pairs with a resolution of 540 × 960. It offers two official evaluation metrics: end-point error (EPE) and n-pixel error (npx). More precisely, EPE represents the average error of disparity estimations, whereas n-pixel error is defined as the percentage of pixels whose disparity estimation errors are greater than n pixels. KITTI 2012 and KITTI 201557,58 are two real-world street-scene benchmarks captured from autonomous driving scenarios. Specifically, KITTI 2012 contains 194 training stereo image pairs and 195 test pairs, while KITTI 2015 comprises 200 training stereo image pairs and 200 test pairs. The two benchmarks share an approximate resolution of 376 × 1240. Sparse ground truth disparity maps are supplied for the training samples, while the testing samples are reserved for online benchmarking with no publicly available ground truth. For KITTI 2012, the official evaluation criterion is the n-pixel error, which aligns with the metric used in the Scene Flow dataset. KITTI 2015 adopts D1 as its official metric, which is defined as the proportion of disparity outliers whose disparity error surpasses both 3 pixels and 5% of the ground-truth disparity. It is worth noting that the evaluation metrics for KITTI 2012/2015 can be calculated either for non-occluded (Noc) regions or for all (All) pixels. Implementation details Our proposed network is implemented using the PyTorch framework, and all training procedures are conducted on an NVIDIA RTX 3090 GPU with the AdamW optimizer (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _1 = 0.9$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _2 = 0.999$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {weight\_decay} = 1.0 \times 10^{-5}$$\end{document}). During preprocessing, training images are randomly cropped to a resolution of 320 × 736, followed by color normalization. The maximum disparity value is configured as 192, and the weighting coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _2$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _3$$\end{document} for the total loss (Eq. 23) are set to 1.0, 0.3, and 0.15, respectively. To facilitate comparisons with SOTA methods on the Scene Flow dataset, the network is trained from scratch for 90 epochs, employing the one-cycle learning rate scheduler59 and the maximum learning rate is set to 0.001. For ablation experiments conducted on the Scene Flow dataset, the network is trained from scratch for 50 epochs to minimize computational cost. For the KITTI 2012/2015 benchmarks, the Scene Flow-pre-trained network is fine-tuned on the combined training set of KITTI 2012/2015 for 500 epochs, adopting the one-cycle learning rate scheduler whose maximum learning rate is set to 0.0002. As shown in Fig. 3 of Section "Encoder-decoder with spatially adaptive modulation (ED-SAM)", ED-SAM exhibits excellent scalability. Thus, we can control the computational cost of the model by tuning the architectural parameters a, b, and c in ED-SAM. Specifically, we first set a=1, b=2, and c=4 in the ED-SAM module to obtain LASM-Net-S, an extremely lightweight model variant. We then scale up the ED-SAM module for cost aggregation on the high-resolution cost volume \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\tilde{C}}_1$$\end{document} by setting a=4, b=8, and c=16, yielding the larger model variant LASM-Net-L. In our comparative experiments, we evaluate the empirical performance of both LASM-Net-S and LASM-Net-L. Ablation studies To evaluate the performance of the proposed method, ablation studies are performed on the Scene Flow dataset using LASM-Net-S. We first construct a baseline model by removing all key components and strategies from LASM-Net-S, followed by gradual integration of them back into the baseline. Specifically, the baseline model is constructed as follows: 1) To verify the effectiveness of multi-scale cost volumes, the baseline model builds only one cost volume at 1/4 resolution; 2) To verify the effectiveness of our spatially adaptive modulation (SAM) module in cost aggregation, the baseline model adopts the same encoder-decoder structure as ED-SAM but excludes the SAM module; 3) To verify the effectiveness of our multi-scale attention-guided fusion (MAGF) strategy, the baseline model fuses the two multi-scale cost volumes through matrix summation after bilinear interpolation for size alignment. In addition, as illustrated in Fig. 5 and Eq. 5, our SAM executes cost volume modulation across all three dimensions (i.e., the attention matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{A}$$\end{document} is in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {R}^{D_{in} \times H_{in} \times W_{in}}$$\end{document}). To validate its necessity, we design a variant module termed SAM-2D, whose attention matrix is defined in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {R}^{H_{in} \times W_{in}}$$\end{document}. The results are summarized in Table 2. As shown in Row 1 of Table 2, the baseline network has fewer parameters, lower MACs, and faster inference speed, but suffers from a noticeable accuracy drop. From Rows 1–7, we observe steady accuracy gains as the proposed modules and strategies are gradually incorporated, while the introduced computational overhead remains relatively low. Notably, a comparison of Row 1 and Row 7 reveals that integrating the multi-scale cost volume strategy, SAM, and MAGF modules reduces the EPE, 1px error, and D1 metrics by 13.89%, 11.91%, and 15.95%, respectively. This improvement comes with only a small overhead of 0.54M additional parameters, 0.39G MACs, and 5 ms extra inference latency. In addition, a comparison of Rows 7 and 8 demonstrates that SAM-2D achieves clearly lower accuracy than SAM. This confirms that modulation along all three dimensions of the cost volume (i.e., disparity, height, and width) is essential for SAM to maintain high precision. In summary, these results clearly validate the effectiveness and computational efficiency of the proposed modules and strategies.Table 2Ablation studies with LASM-Net-S on the Scene Flow dataset. Abbreviations: SAM, Spatially Adaptive Modulation; MAGF, Multi-scale Attention-Guided Fusion. † and * denote the baseline network and our full LASM-Net-S model, respectively.No.Volume ScalesModulationVolume FusionEPE1pxD1Param.MACsTime1/81/4(pixel)(%)(%)(M)(G)(ms)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {1}^\dagger$$\end{document}1/4×××0.7278.2232.7463.3422.171421/4,1/8××Sum0.6707.9002.5043.7222.721831/4,1/8SAM×Sum0.6387.4982.3773.7722.781841/4,1/8×SAMSum0.6417.4552.3543.8223.261951/4,1/8××MAGF0.6547.6972.4553.7322.951861/4,1/8SAMSAMSum0.6347.3012.3253.8723.3219\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {7}^*$$\end{document}1/4,1/8SAMSAMMAGF0.6267.2432.3083.8823.561981/4,1/8SAM-2DSAM-2DMAGF0.6547.7022.4683.8223.3219Table 3Comparisons on the Scene Flow test set. Different models are categorized into accuracy-oriented and speed-oriented categories. * denotes the data tested with the open-source code. The best and second-best performance in each category are shown in bold font and underlined styling.ModelEPE (pixel)Params. (M)MACs (G)Time (ms)GPU TypeAccuracy-orientedPSMNet351.095.21084.21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${295}^*$$\end{document}RTX 3090CAR-Net600.923.1-117RTX 2080TiDeepPruner-Best610.867.4515.62\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$133^*$$\end{document}RTX 30903D-MSNet200.801.8615.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$512^*$$\end{document}RTX 3090GANet-deep360.786.6-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1760^*$$\end{document}RTX 3090GwcNet-gc320.776.81039.53\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$203^*$$\end{document}RTX 3090UPFNet620.71--270RTX 3090ACVNet130.48--200RTX 3090speed-orientedADCPNet-LN9301.433.6---FBPGNet631.201.3-44RTX 30902D-MSNet201.142.2136.89\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$103^*$$\end{document}RTX 3090BGNet241.173.057.4625RTX 2080TiDeepPruner-Fast610.977.5219.13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$48^*$$\end{document}RTX 3090AANet500.873.9180.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$65^*$$\end{document}RTX 3090CKDNet640.872.2-96Titan XpSADSNet-L-N7310.901.536.8913RTX 3090GFANet650.82--63RTX 3090ACA-Net660.774.3199.4487RTX 3090SADSNet-L-N7-ref310.731.657.0124RTX 3090LightStereo-S230.733.422.7117RTX 3090HybridNet670.721.1171.34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$93^*$$\end{document}RTX 3090CoEx180.692.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$53.58^*$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$22^*$$\end{document}RTX 3090Fast-ACVNet130.64--39RTX 3090SAGU-Net-fr260.635.5184.4741RTX 3090LightStereo-L230.5924.391.8537RTX 3090LASM-Net-S (Ours)0.633.923.5619RTX 3090LASM-Net-L (Ours)0.588.137.2125RTX 3090 Table 2Ablation studies with LASM-Net-S on the Scene Flow dataset. Abbreviations: SAM, Spatially Adaptive Modulation; MAGF, Multi-scale Attention-Guided Fusion. † and * denote the baseline network and our full LASM-Net-S model, respectively.No.Volume ScalesModulationVolume FusionEPE1pxD1Param.MACsTime1/81/4(pixel)(%)(%)(M)(G)(ms)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {1}^\dagger$$\end{document}1/4×××0.7278.2232.7463.3422.171421/4,1/8××Sum0.6707.9002.5043.7222.721831/4,1/8SAM×Sum0.6387.4982.3773.7722.781841/4,1/8×SAMSum0.6417.4552.3543.8223.261951/4,1/8××MAGF0.6547.6972.4553.7322.951861/4,1/8SAMSAMSum0.6347.3012.3253.8723.3219\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {7}^*$$\end{document}1/4,1/8SAMSAMMAGF0.6267.2432.3083.8823.561981/4,1/8SAM-2DSAM-2DMAGF0.6547.7022.4683.8223.3219 Ablation studies with LASM-Net-S on the Scene Flow dataset. Abbreviations: SAM, Spatially Adaptive Modulation; MAGF, Multi-scale Attention-Guided Fusion. † and * denote the baseline network and our full LASM-Net-S model, respectively. Table 3Comparisons on the Scene Flow test set. Different models are categorized into accuracy-oriented and speed-oriented categories. * denotes the data tested with the open-source code. The best and second-best performance in each category are shown in bold font and underlined styling.ModelEPE (pixel)Params. (M)MACs (G)Time (ms)GPU TypeAccuracy-orientedPSMNet351.095.21084.21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${295}^*$$\end{document}RTX 3090CAR-Net600.923.1-117RTX 2080TiDeepPruner-Best610.867.4515.62\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$133^*$$\end{document}RTX 30903D-MSNet200.801.8615.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$512^*$$\end{document}RTX 3090GANet-deep360.786.6-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1760^*$$\end{document}RTX 3090GwcNet-gc320.776.81039.53\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$203^*$$\end{document}RTX 3090UPFNet620.71--270RTX 3090ACVNet130.48--200RTX 3090speed-orientedADCPNet-LN9301.433.6---FBPGNet631.201.3-44RTX 30902D-MSNet201.142.2136.89\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$103^*$$\end{document}RTX 3090BGNet241.173.057.4625RTX 2080TiDeepPruner-Fast610.977.5219.13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$48^*$$\end{document}RTX 3090AANet500.873.9180.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$65^*$$\end{document}RTX 3090CKDNet640.872.2-96Titan XpSADSNet-L-N7310.901.536.8913RTX 3090GFANet650.82--63RTX 3090ACA-Net660.774.3199.4487RTX 3090SADSNet-L-N7-ref310.731.657.0124RTX 3090LightStereo-S230.733.422.7117RTX 3090HybridNet670.721.1171.34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$93^*$$\end{document}RTX 3090CoEx180.692.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$53.58^*$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$22^*$$\end{document}RTX 3090Fast-ACVNet130.64--39RTX 3090SAGU-Net-fr260.635.5184.4741RTX 3090LightStereo-L230.5924.391.8537RTX 3090LASM-Net-S (Ours)0.633.923.5619RTX 3090LASM-Net-L (Ours)0.588.137.2125RTX 3090 Comparisons on the Scene Flow test set. Different models are categorized into accuracy-oriented and speed-oriented categories. * denotes the data tested with the open-source code. The best and second-best performance in each category are shown in bold font and underlined styling. Comparisons on scene flow Table 3 compares the proposed LASM-Net with SOTA stereo matching models published in recent years on the Scene Flow test set. For more intuitive insights, we also plot the performance of each network in an “EPE vs. Frame Rate” 2D coordinate system in Fig. 7. As can be seen in Table 3, compared to Fast-ACVNet, SAGU-Net-fr, and LightStereo-L—three state-of-the-art speed-oriented networks that achieve the highest accuracy—our LASM-Net-S delivers comparable accuracy (EPE values: 0.63 vs. 0.64, 0.63, and 0.59) while achieving significantly faster inference speed (frame rate: 52.6 vs. 25.6, 24.4, and 27.0 FPS). Furthermore, our LASM-Net-L achieves a remarkable EPE value of 0.58, which not only enables it to outperform all other speed-oriented methods in accuracy but also maintains real-time inference with a frame rate of up to 40 FPS. Even when compared to the accuracy-oriented models, our LASM-Net-S and LASM-Net-L achieve accuracy that is notably higher than that of most counterparts (with the exception of ACVNet), while achieving substantially higher inference speeds than all accuracy-oriented models. Fig. 7 further demonstrates that our models lie closer to the bottom-right corner, showcasing a favorable trade-off between accuracy and computational efficiency. Additionally, Fig. 8 presents qualitative visualization results on the Scene Flow test set. As can be seen, our LASM-Net generates more precise disparity predictions in textureless regions, depth-discontinuity regions, and fine-grained structural details. A closer inspection of these visual examples shows that the advantage of LASM-Net is mainly reflected in regions where reliable matching requires both local structural details and broader contextual cues. In textureless regions, the low-resolution contextual information introduced by MAGF helps suppress locally ambiguous responses. Around depth discontinuities and thin structures, the high-resolution cost volume and feature-guided modulation preserve sharper disparity transitions. These observations are consistent with the motivation of combining ED-SAM and MAGF for lightweight yet accurate cost aggregation.Fig. 7Comparison of end-point error and frame rate between different methods on the Scene Flow test set.Fig. 8Visual comparisons of disparity prediction outputs on the Scene Flow test set. The end-point error of each predicted disparity map is annotated in its upper-right corner.Table 4Results on KITTI 2012/2015 benchmarks. Lower is better for each metric. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} indicates the runtime we tested using open-source code on our hardware. Existing models in the table are listed in descending order of inference time. In each category (accuracy-oriented and speed-oriented models), the optimal and suboptimal results are presented in bold font and with underlining, respectively.ModelKITTI 2012KITTI 2015Time (ms)3px-Noc(%)3px-All(%)D1-Noc(%)D1-All(%)Accuracy-orientedGANet-15361.361.801.731.931500CSPN681.191.531.611.741000AcfNet691.171.541.721.894803D-MSNet20--1.922.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$468^*$$\end{document}AAED-Net70--1.892.09280Waveletstereo711.662.182.422.65270UPFNet621.121.461.501.62250ACVNet371.131.47-1.65200GwcNet-gc321.321.701.922.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$185^*$$\end{document}CFNet101.231.581.731.88180CAR-Net601.541.962.212.36110DPCTF-S721.311.72-1.98110Speed-oriented2D-MSNet20--2.542.83\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$97^*$$\end{document}MFPSNet73--2.983.2787HybridNet671.762.23-2.42\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$86^*$$\end{document}AANet501.912.422.322.55\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$57^*$$\end{document}FADNet252.042.462.392.6050DeepPruner-Fast612.262.732.352.59\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$40^*$$\end{document}FBPGNet63--4.044.3040ADCPNet-L+refine302.202.662.522.7127RST-Stereo74---2.8422ADCPNet-LN9302.443.042.843.0921SADSNet-LN7+refine312.012.552.312.5421\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {RTS}^2$$\end{document}Net29--3.223.5620MADNet285.346.394.274.6620RTSNet752.432.90-3.4120LightStereo-S231.882.342.112.3017P3SNet+513.554.424.324.7215RTSMNet19--3.573.8814AnyNet_C32273.604.334.634.9513P3SNet513.654.464.655.0512SADSNet-L-N7312.122.662.532.7412CMNet22-3.41-3.848ADCPNet-MN7303.143.843.693.987HRSNet76---4.287GWDRNet-S772.653.173.373.646LASM-Net-S (Ours)1.702.112.032.2319LASM-Net-L (Ours)1.491.831.871.9925 Comparison of end-point error and frame rate between different methods on the Scene Flow test set. Visual comparisons of disparity prediction outputs on the Scene Flow test set. The end-point error of each predicted disparity map is annotated in its upper-right corner. Table 4Results on KITTI 2012/2015 benchmarks. Lower is better for each metric. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} indicates the runtime we tested using open-source code on our hardware. Existing models in the table are listed in descending order of inference time. In each category (accuracy-oriented and speed-oriented models), the optimal and suboptimal results are presented in bold font and with underlining, respectively.ModelKITTI 2012KITTI 2015Time (ms)3px-Noc(%)3px-All(%)D1-Noc(%)D1-All(%)Accuracy-orientedGANet-15361.361.801.731.931500CSPN681.191.531.611.741000AcfNet691.171.541.721.894803D-MSNet20--1.922.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$468^*$$\end{document}AAED-Net70--1.892.09280Waveletstereo711.662.182.422.65270UPFNet621.121.461.501.62250ACVNet371.131.47-1.65200GwcNet-gc321.321.701.922.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$185^*$$\end{document}CFNet101.231.581.731.88180CAR-Net601.541.962.212.36110DPCTF-S721.311.72-1.98110Speed-oriented2D-MSNet20--2.542.83\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$97^*$$\end{document}MFPSNet73--2.983.2787HybridNet671.762.23-2.42\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$86^*$$\end{document}AANet501.912.422.322.55\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$57^*$$\end{document}FADNet252.042.462.392.6050DeepPruner-Fast612.262.732.352.59\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$40^*$$\end{document}FBPGNet63--4.044.3040ADCPNet-L+refine302.202.662.522.7127RST-Stereo74---2.8422ADCPNet-LN9302.443.042.843.0921SADSNet-LN7+refine312.012.552.312.5421\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {RTS}^2$$\end{document}Net29--3.223.5620MADNet285.346.394.274.6620RTSNet752.432.90-3.4120LightStereo-S231.882.342.112.3017P3SNet+513.554.424.324.7215RTSMNet19--3.573.8814AnyNet_C32273.604.334.634.9513P3SNet513.654.464.655.0512SADSNet-L-N7312.122.662.532.7412CMNet22-3.41-3.848ADCPNet-MN7303.143.843.693.987HRSNet76---4.287GWDRNet-S772.653.173.373.646LASM-Net-S (Ours)1.702.112.032.2319LASM-Net-L (Ours)1.491.831.871.9925 Results on KITTI 2012/2015 benchmarks. Lower is better for each metric. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^*$$\end{document} indicates the runtime we tested using open-source code on our hardware. Existing models in the table are listed in descending order of inference time. In each category (accuracy-oriented and speed-oriented models), the optimal and suboptimal results are presented in bold font and with underlining, respectively. Comparisons on KITTI 2012 & KITTI 2015 The widely used KITTI 2012/2015 benchmarks facilitate comparisons between the proposed LASM-Net and accuracy-oriented as well as speed-oriented models. Results are summarized in Table 4. Additionally, for more intuitive insights, we plot the performance of the compared models on the KITTI 2012 and KITTI 2015 datasets in an “Error Rate vs. Frame Rate” 2D coordinate system, as shown in Fig. 9 and Fig. 10, respectively. Table 4 shows that the proposed LASM-Net achieves the best accuracy among all speed-oriented models. Specifically, the inference speed of ADCPNet-LN9, SADSNet-LN7+refine, LightStereo-S, and our LASM-Net-S is comparable, but the accuracy of our LASM-Net-S on the KITTI 2012 dataset is 30.6%, 17.3%, and 9.8% higher than that of these three models regarding the 3px-All metric. On the KITTI 2015 dataset, its accuracy is 27.8%, 12.2%, and 3.0% higher regarding the D1-All metric. In addition, our LASM-Net-L achieves accuracy comparable to that of many SOTA accuracy-oriented models while maintaining substantially higher inference speed. On the KITTI 2015 dataset, our LASM-Net-L outperforms five out of twelve accuracy-oriented models on the D1-All metric.Fig. 9Error rate (3px-All) versus frame rate on KITTI 2012 dataset.Fig. 10Error rate (D1-All) versus frame rate on KITTI 2015 dataset.Fig. 11Visual comparison results on the KITTI 2012 dataset. Disparity maps predicted by the compared models and their corresponding error maps are provided, where lighter colors indicate larger errors. The 3px-All error rate is annotated below each disparity map. Our proposed LASM-Net-L produces sharper disparity edges at object boundaries while delivering more precise disparity estimations in flat, textureless regions. Error rate (3px-All) versus frame rate on KITTI 2012 dataset. Error rate (D1-All) versus frame rate on KITTI 2015 dataset. Visual comparison results on the KITTI 2012 dataset. Disparity maps predicted by the compared models and their corresponding error maps are provided, where lighter colors indicate larger errors. The 3px-All error rate is annotated below each disparity map. Our proposed LASM-Net-L produces sharper disparity edges at object boundaries while delivering more precise disparity estimations in flat, textureless regions. Figs. 9 and 10 provide an intuitive illustration of the performance of each model. As can be seen from these figures, the proposed LASM-Net-S/L models are distributed closer to the bottom-right corner, indicating that they achieve a better trade-off between computational overhead and disparity accuracy. Figure 11 shows the visual comparison results on the KITTI 2012 dataset. As observed, the proposed LASM-Net-L yields sharper edges at object boundaries while delivering precise disparity estimations in flat, textureless regions. This qualitative improvement explains why LASM-Net achieves favorable quantitative results among speed-oriented methods. The proposed spatially adaptive modulation helps reweight unreliable matching responses according to image-dependent cues, while MAGF adaptively injects contextual information from the low-resolution cost volume into high-resolution matching. Nevertheless, mismatches may still occur in extremely ambiguous regions, such as repetitive patterns, reflective surfaces, very thin objects, or regions affected by severe image degradation. Robustness evaluation In this subsection, we analyze the robustness of the proposed model under challenging conditions, including illumination variation, occlusion, and noise interference. Specifically, the LASM-Net-S model is first trained on the Scene Flow training set without these kinds of disturbances. Subsequently, we impose disturbances of different types and intensities on the images from the Scene Flow test set, and then evaluate the performance variation of the LASM-Net-S model on this corrupted test set. Firstly, we randomly adjust the illumination intensity of the left and right images of all stereo pairs in the Scene Flow test set to investigate the model’s performance change. Specifically, the images with illumination variation are defined as:25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{I}_{L,\text {illum}} = \boldsymbol{I}_{L} \times \text {rand}(1-r,1+r),\quad \boldsymbol{I}_{R,\text {illum}} = \boldsymbol{I}_{R} \times \text {rand}(1-r,1+r) \end{aligned}$$\end{document}where r denotes the variation factor controlling the intensity of illumination perturbations, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {rand}(1-r,1+r)$$\end{document} means randomly sampling a scalar from the interval [1-r,1+r]. Figure 12a illustrates the performance of LASM-Net-S under different levels of illumination variation on the Scene Flow test set. It can be observed that the model’s EPE remains nearly unchanged across r ∈ [0, 0.3], and only exhibits a slight upward trend when r exceeds 0.3. Overall, the performance degradation under illumination variations is negligible, indicating that our method is highly robust to illumination changes and maintains stable matching accuracy under diverse lighting conditions. Secondly, we evaluate the robustness of the model against occluded regions. We adopt a random mean occlusion strategy on the Scene Flow test set to evaluate the robustness of the proposed model. As shown in Algorithm 1, given the original left and right images \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{I}_{L}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{I}_{R}$$\end{document}, we first randomly generate the position and size of an occlusion block within the predefined maximum size t. Then, we extract the local image patch corresponding to the occlusion area and calculate its channel-wise mean value. Then, we replace the original pixel values of the occlusion area with the computed mean value. This operation is independently applied to both the left and right stereo images to simulate occlusions. As shown in Fig. 12b, even when the maximum occlusion size t increases from 0 to 100 pixels, the End-Point Error (EPE) of LASM-Net-S remains consistently low and stable, with only a marginal increase at the highest occlusion level. This result demonstrates that the proposed method maintains high matching accuracy under varying degrees of occlusion, verifying its strong robustness to occluded regions in stereo matching tasks.Fig. 12Robustness evaluation of LASM-Net-S on the Scene Flow test set: (a) Illumination variation test; (b) Occlusion test. Robustness evaluation of LASM-Net-S on the Scene Flow test set: (a) Illumination variation test; (b) Occlusion test. Finally, Gaussian white noise with a standard deviation of σ is added to all images in the Scene Flow test set to test the performance variation of LASM-Net-S. Specifically, given the original stereo image pair \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{I}_{L}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{I}_{R}$$\end{document}, the noisy stereo image pair can be formulated as:26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{I}_{L,\text {noise}} = \boldsymbol{I}_{L} + \text {randn}(0,\sigma ),\quad \boldsymbol{I}_{R,\text {noise}} = \boldsymbol{I}_{R} + \text {randn}(0,\sigma ) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {randn}(0,\sigma )$$\end{document} denotes the Gaussian white noise matrix with the same dimension as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{I}_{L/R}$$\end{document}. We point out that the images in the Scene Flow training set are relatively clean and do not contain obvious noise. Consequently, as shown in Fig. 13, LASM-Net-S trained on the original Scene Flow training set is sensitive to Gaussian noise intensity, and high-intensity Gaussian noise leads to a large end-point error. This observation reveals a limitation of the proposed model when it is trained only on clean images: severe sensor noise or strong image degradation may substantially disturb the learned matching cost distribution. Therefore, noise augmentation or domain-specific fine-tuning is necessary when LASM-Net is deployed in scenarios with heavy sensor noise or severe image corruption. To address extreme scenarios with high-intensity noise, we introduce Gaussian white noise augmentation into the training process and retrain the LASM-Net-S model. Figure 13 presents the performance of the retrained model on the Scene Flow test set corrupted by noise with different intensity levels. It can be observed that the model trained with noise augmentation achieves outstanding robustness against noise interference. Algorithm 1Random mean occlusion augmentation for stereo images. Random mean occlusion augmentation for stereo images. Fig. 13Robustness evaluation of LASM-Net-S against Gaussian noise on the Scene Flow test set. The green line corresponds to our LASM-Net-S model trained on the original Scene Flow training dataset, which exhibits a sharp increase in end-point error (EPE) as the noise intensity rises. The orange line represents our LASM-Net-S model retrained with Gaussian noise augmentation, maintaining consistently stable EPE across all noise levels. σ denotes the standard deviation of the added Gaussian noise. Robustness evaluation of LASM-Net-S against Gaussian noise on the Scene Flow test set. The green line corresponds to our LASM-Net-S model trained on the original Scene Flow training dataset, which exhibits a sharp increase in end-point error (EPE) as the noise intensity rises. The orange line represents our LASM-Net-S model retrained with Gaussian noise augmentation, maintaining consistently stable EPE across all noise levels. σ denotes the standard deviation of the added Gaussian noise. Generalization across datasets To verify the generalization ability of the proposed LASM-Net-L model, we employ the LASM-Net-L model pre-trained on the Scene Flow dataset and directly test its performance on the training stereo images of the Middlebury dataset78. Figure 14 intuitively presents the visualization results of predicted disparity maps. It is worth pointing out that our model is primarily designed to yield more precise stereo matching results at a lower computational cost. Specifically, we prioritize a desirable balance between computational complexity and disparity accuracy instead of optimizing for stronger generalization capability. Nevertheless, Fig. 14 indicates that the proposed method still achieves decent generalization ability.Fig. 14Generalization results of the proposed LASM-Net-L model on images of the Middlebury dataset, where LASM-Net-L is trained only on the Scene Flow dataset and tested on training images of the Middlebury dataset. Generalization results of the proposed LASM-Net-L model on images of the Middlebury dataset, where LASM-Net-L is trained only on the Scene Flow dataset and tested on training images of the Middlebury dataset. Deployment on edge computing devices To evaluate the performance of the proposed model on embedded AI hardware platforms with limited computational capability, we deploy and test the proposed LASM-Net on NVIDIA Jetson TX2, a commonly used embedded AI hardware platform. The detailed results are presented in Table 5. Meanwhile, we plot the metrics of D1-All and frame rate on the KITTI 2015 dataset as a figure to provide a more intuitive comparison, as shown in Fig. 15. As shown in the table and figure, the proposed model achieves a better trade-off between accuracy and inference speed. Meanwhile, in practical applications, its model scale can be easily configured according to the available computational resources, enabling it to accommodate diverse computational scenarios.Table 5Test results on the NVIDIA Jetson TX2 hardware platform. Bold and underlined values indicate the best and second-best results in each column, respectively.ModelKITTI 2012KITTI 2015Frame rate (FPS)3px-All (%)D1-All (%)DeepPruner-Fast612.732.590.68MADNet286.394.663.85AnyNet-C1276.427.1010.40AnyNet-C32274.334.953.50\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {RTS}^2$$\end{document}Net29–3.562.30ADCPNet-MN7303.843.988.93ADCPNet-LN9303.043.092.49ADCPNet-L-Refine302.662.711.89LBSM-L792.772.814.45LBSM-L-Refine792.412.522.26LASM-Net-S (Ours)2.112.234.18LASM-Net-L (Ours)1.831.992.01 Table 5Test results on the NVIDIA Jetson TX2 hardware platform. Bold and underlined values indicate the best and second-best results in each column, respectively.ModelKITTI 2012KITTI 2015Frame rate (FPS)3px-All (%)D1-All (%)DeepPruner-Fast612.732.590.68MADNet286.394.663.85AnyNet-C1276.427.1010.40AnyNet-C32274.334.953.50\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {RTS}^2$$\end{document}Net29–3.562.30ADCPNet-MN7303.843.988.93ADCPNet-LN9303.043.092.49ADCPNet-L-Refine302.662.711.89LBSM-L792.772.814.45LBSM-L-Refine792.412.522.26LASM-Net-S (Ours)2.112.234.18LASM-Net-L (Ours)1.831.992.01 Test results on the NVIDIA Jetson TX2 hardware platform. Bold and underlined values indicate the best and second-best results in each column, respectively. Fig. 15Error rate (D1-All) and Frame rate comparisons tested with NVIDIA Jetson TX2 on the KITTI 2015 dataset. Error rate (D1-All) and Frame rate comparisons tested with NVIDIA Jetson TX2 on the KITTI 2015 dataset. Limitations and future work It should be pointed out that the proposed model still has several limitations. First, the goal of our LASM-Net is not to outperform all accuracy-oriented stereo matching models in terms of absolute disparity accuracy. Some models based on 4D cost volumes, 3D convolutions, iterative refinement, or transformer-based interaction can still achieve stronger accuracy, but usually require substantially higher computational cost. Therefore, our LASM-Net is more suitable for scenarios where both accuracy and real-time efficiency are important. Second, the robustness evaluation in Section “Robustness evaluation” shows that LASM-Net-S trained only on clean Scene Flow images is sensitive to high-intensity Gaussian noise. Although noise augmentation significantly improves robustness, this result indicates that domain-specific augmentation or fine-tuning may be necessary when deploying the model in sensor-noisy environments. Third, although our model achieves an impressive accuracy-efficiency trade-off, the inference speed of the proposed model is slower than that of the most lightweight existing models. Further improving its inference speed by introducing neural network inference acceleration techniques (e.g., NVIDIA TensorRT) will be one of the major focuses of our future work. Conclusion This study addresses the critical trade-off between accuracy and computational overhead in stereo matching, a long-standing challenge that limits the practical deployment of high-accuracy models in resource-constrained or low-latency scenarios. Existing speed-oriented stereo matching models, while improving inference efficiency, suffer from significant accuracy degradation, resulting in a notable performance gap compared to accuracy-oriented counterparts. To bridge this gap, a lightweight stereo matching model LASM-Net is designed to achieve high disparity accuracy with low computational burden. The key contributions of this work are summarized as follows: First, we abandon the computationally expensive 4D cost volume aggregation with 3D convolutions, instead adopting multi-scale 3D cost volumes combined with pure 2D operations to reduce computational burden. To mitigate potential accuracy loss, we introduce spatially adaptive modulation matrices to modulate the 3D cost volume across all three dimensions, enabling effective exploitation of disparity distribution cues and spatial contextual information. Second, to resolve the dilemma where speed-oriented models sacrifice receptive field size and contextual information due to reduced convolutional layers, we proposed a multi-scale attention-guided fusion strategy. This strategy compensates for the missing contextual information using low-resolution cost volumes, ensuring sufficient feature representation without increasing computational cost. Comprehensive evaluations validate that the integration of these core designs and other dedicated optimizations enables LASM-Net to achieve superior performance in terms of accuracy and running speed. Specifically, LASM-Net effectively narrows the accuracy gap between lightweight and accuracy-oriented stereo matching models, verifying the feasibility of balancing high efficiency and high accuracy in stereo matching networks. Future work will further investigate the acceleration potential of neural network optimization techniques, including model pruning, reparameterization, and sparsification, on the proposed model.