Fast Feature Matching of UAV Images via Matrix Band Reduction-based GPU Data Schedule
San Jiang, Kan You, Wanshou Jiang, Qingquan Li · 2025
Read Fast Feature Matching of UAV Images via Matrix Band Reduction-based GPU Data Schedule by San Jiang, Kan You, Wanshou Jiang, Qingquan Li (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-22089.txt.
Introduction
San Jiang and colleagues speed up the matching of features across drone photos, a step that dominates the cost of building three-dimensional models from images. Their method, from twenty twenty-five, first picks likely image pairs with a fast retrieval search. It then groups images into compact blocks using a matrix reordering trick that follows image connections, so the graphics card wastes less time loading data. Matching runs inside each block with an accelerated hashing scheme, and rough matches are cleaned with local geometry checks plus a standard robust fitting step.
On large drone data sets, the schedule ran far faster than tree-based matching, with speed gains reported in the dozens of times. Accuracy in later camera and model adjustments stayed comparable. The authors offer the approach for survey teams facing very large aerial photo collections.
Catalog record
| CREATOR | San Jiang, Kan You, Wanshou Jiang, Qingquan Li |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-22089
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Fast Feature Matching of UAV Images via Matrix Band Reduction-based GPU Data Schedule San Jiang, Kan You, Wanshou Jiang, Qingquan Li Abstract Feature matching dominats the time costs in structure from motion (SfM). The primary contribution of this study is a GPU data schedule algorithm for efficient feature matching of Unmanned aerial vehicle (UAV) images. The core idea is to divide the whole dataset into blocks based on the matrix band reduction (MBR) and achieve efficient feature matching via GPU-accelerated cascade hashing. First, match pairs are selected by using an image retrieval technique, which converts images into global descriptors and searches high-dimension nearest neighbors with graph indexing. Second, compact image blocks are iteratively generated from a MBR-based data schedule strategy, which exploits image connections to avoid redundant data IO (input/output) burden and increases the usage of GPU computing power. Third, guided by the generated image blocks, feature matching is executed sequentially within the framework of GPU-accelerated cascade hashing, and initial candidate matches are refined by combining a local geometric constraint and RANSAC-based global verification. For further performance improvement, these two seps are designed to execute parallelly in GPU and CPU. Finally, the performance of the proposed solution is evaluated by using large-scale UAV datasets. The results demonstrate that it increases the efficiency of feature matching with speedup ratios ranging from 77.0 to 100.0 compared with KD-Tree based matching methods, and achieves comparable accuracy in relative and absolute bundle adjustment (BA). The proposed algorithm is an efficient solution for feature matching of UAV images. 1 Introduction Unmanned aerial vehicle (UAV) has been increasingly becoming a critical remote sensing platform in photogrammetry and computer vision ( Jiang et al. 2021b ; Yao et al. 2019 ) , and it has been widely used in a range of applications, e.g., indoor inspection of transmission lines ( Jiang et al. 2017 ) , procedural 3D modeling of urban facilities ( Wang et al. 2023 ) , and precision management of agricultural plants ( Zheng et al. 2021 ) . The precise calculation of camera poses is an essential preliminary step, which has been usually implemented through structure from motion (SfM) ( Jiang et al. 2020a ) . In the workflow of SfM, the time costs are dominated by feature matching due to the quadratic combination complexity of match pairs and nearest neighbor searching (NNS) of high-dimension feature descriptors. Thus, fast feature matching is required for recent SfM systems. In the literature, extensive work has been reported to address these issues. To decrease the combination complexity of match pairs, auxiliary data and visual similarity are the widely used clues to select overlapped image pairs and avoid exhaustive feature matching ( Hartmann et al. 2016 ) , e.g., the onboard POS (Positioning and Orientation System) data from UAV platforms ( Jiang and Jiang 2020 ; Remondino and Gerke 2015 ) and the geometric priors from data acquisitions ( Jiang et al. 2024 ; Schönberger et al. 2014 ) . In addition, to cope with oblique data acquisition and geometry-aware trajectory planning ( Li et al. 2023 ) , visual similarity has also been extensively exploited in modern SfM systems, which usually use pre-trained visual words to encode images into fix-length vectors and convert the problem of selecting overlapped image pairs into the one of nearest neighbor searching among the vectors, such as BoW (Bag-of-Words) ( Nister and Stewenius 2006 ; Havlena and Schindler 2014 ; Jiang et al. 2022a ) and VLAD (Vector of Locally Aggregated Descriptors) ( Hou et al. 2023 ; Jiang et al. 2023 ) . By using these techniques, the time complexity of feature matching is significantly decreased, which is transited from a quadratic relationship to a linear one with respect to the number of involved UAV images. Match pair selection reduces the combination complexity of images. However, high time costs are still required in the exhaustive nearest neighbor searching among two sets of high-dimension feature descriptors ( Hartmann et al. 2016 ) , especially for the widely used KD-Tree based SIFT feature matching ( Lowe 2004 ) . To accelerate feature matching, existing methods can be divided into two groups, i.e., redesign of feature descriptors and acceleration of feature matching. For the former, feature descriptors are designed to decrease the computational burden, such as the ORB ( Rublee et al. 2011 ) with binary descriptors. However, as verified in recent work, SIFT-based feature matching is still the most robust algorithm in practical engineering applications ( Ji et al. 2023 ; Jiang et al. 2021a ) . For the latter, the hardware acceleration technique has been exploited to speed up feature matching. The classical examples include the SIFTGPU ( Wu 2011 ) and popsift ( Griwodz et al. 2018 ) algorithms that exploit GPU (Graphic Processing Unit) for feature matching. In contrast to the KD-Tree based approximate nearest neighbor (ANN) searching, hashing-based methods have also gained lots of attention in feature matching, whose core idea is to convert the real-value descriptors to discrete binary codes via hashing functions and map similar descriptors of query and database images into the same cell ( Cheng et al. 2014 ) . To achieve higher speedup ratios, GPU acceleration and data schedule have been integrated into hashing-based matching pipelines in the recent work ( Xu et al. 2017 ; Zhang et al. 2023 ) , which use hardware acceleration for calculating hash codes and similarity scores and exploit data schedule to relief input/output (IO) burdens. Compared with classical SIFT matching algorithms, hashing-based methods can achieve promising speedup ratios for feature matching of UAV images. However, some issues still exist in the proposed solutions. First, the performance of existing data schedule depends on the connection structures between images. In the work of Xu et al. 2017 , the group-block based data schedule strategy assumes dense image connections. For oblique photogrammetry, image connections are usually very sparse, and it degenerates the overall performance of GPU-based hashing matching. Second, existing data schedule strategies emphasize more on the IO burdens, however, ignoring the parallel computing power of modern GPU. For the data schedule used in Zhang et al. 2023 , image loading and free sequences are first generated based on the image connection structure, and feature matching is executed sequentially according to loading and free sequence. In other words, only one image is matched against other images in GPU memory. Third, outliers are prone to retain in feature matching due to the usage of binary hashing codes for calculating descriptor similarity scores. It would cause high time cost in classical RANSAC (Random Sample Consensus) based outlier removal ( Lu et al. 2016 ) . Thus, this study proposes a matrix band reduction-based GPU data schedule for fast feature matching of UAV images. Our main contributions are summarized as follows: (1) we propose a matrix band reduction (MBR) based data schedule strategy to divide the sparsely connected view graph into compact blocks, which can adapt well to the connection structure of images and the memory volume of GPUs; (2) we design a cascade hashing based feature matching workflow that integrates the spatial angular order (SAO) based local constraint and RANSAC-based global verification for outlier removal. Combined with MBR-based data schedule, the