Point Cloud Quality Assessment Using the Perceptual Clustering Weighted Graph (PCW-Graph) and Attention Fusion Network
Abdelouahed Laazoufi a, ,Mohammed El Hassouni a and Hocine Cherifi b · 2025
Read Point Cloud Quality Assessment Using the Perceptual Clustering Weighted Graph (PCW-Graph) and Attention Fusion Network by Abdelouahed Laazoufi a, ,Mohammed El Hassouni a and Hocine Cherifi b (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-2506-04081.txt.
Introduction
Abdelouahed Laazoufi, Mohammed El Hassouni, and Hocine Cherifi rate the quality of 3D point clouds without a reference copy, in a 2025 paper. Point clouds drive virtual reality and scanning, and judging their quality without the original is hard. Their method builds a weighted graph that clusters points by how perception groups them. An attention fusion network then weighs the graph's signals into one quality score. The abstract is brief, so details stay high-level.
Catalog record
| CREATOR | Abdelouahed Laazoufi a, ,Mohammed El Hassouni a and Hocine Cherifi b |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 64 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-04081
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Point Cloud Quality Assessment Using the Perceptual Clustering Weighted Graph (PCW-Graph) and Attention Fusion Network Abdelouahed Laazoufi a, ,Mohammed El Hassouni a and Hocine Cherifi b Abstract No-Reference Point Cloud Quality Assessment (NR-PCQA) is critical for evaluating 3D content in real-world applications where reference models are unavailable. Existing NR-PCQA methods often fail to capture structural and perceptual relationships, leading to inconsistent quality predictions. To address this issue, we propose a Perceptual Clustering Weighted Graph (PCW-Graph) that models point cloud distortions through a structured learning process. Our approach follows three main steps: (1) Perceptual Feature Extraction - Color, curvature, and saliency features are computed to capture geometric and visual distortions. (2) Graph-Based Representation — Perceptual clusters form nodes, and weighted edges encode feature similarities, ensuring spatial and structural consistency. (3) Adaptive Quality Prediction - A Graph Attention Fusion Network (GAF) dynamically refines feature importance and estimates quality scores via regression to improve correlation with human perception. We evaluate the proposed method on three benchmark datasets: WPC, SJTU PCQA, and ICIP2020. It achieves high Pearson (PLCC ≥ 0.93 \geq 0.93 ) and Spearman (SRCC ≥ 0.91 \geq 0.91 ) correlation coefficients. These results show better alignment with human perceptual scores than alternative NR-PCQA methods. The approach also reduces Root Mean Square Error (RMSE) by up to 15%. This confirms its effectiveness and robustness for blind point cloud quality assessment. 1 Introduction The adoption of 3D models has recently grown significantly across multiple fields, such as architecture, cultural heritage preservation, virtual and mixed reality, and computer-aided diagnostics. However, processes like compression and simplification applied to these 3D models can introduce distortions that degrade the visual quality of 3D point clouds. This issue has increased the need for effective methods to evaluate perceived quality. In the past, distortion levels in 3D models were assessed by human observers, a process that is time-consuming and resource-intensive. To improve efficiency, objective methods have been introduced as a solution [ 1 ] . These methods employ automatic measurements that simulate the decisions made by an optimal human observer. These metrics can be categorized into three types: Full Reference (FR) [ 2 , 3 , 4 , 5 , 6 ] , Reduced Reference (RR) [ 7 , 8 ] , and No Reference (NR) [ 9 , 10 , 11 , 12 , 13 , 14 , 15 ] . Among these, blind methods, which do not rely on reference models, are gaining increasing importance in real-world applications [ 16 , 17 , 18 , 19 ] . A 3D point cloud (PC) consists of a collection of points defined by geometric coordinates and potentially other attributes such as color, reflectance, or surface normals. Three-dimensional point clouds contain information about the shape and geometry of an object, with points irregularly arranged in space compared to images and videos. Therefore, effective methods for feature extraction from these scattered points must be explored to assess quality accurately. Currently, very few metrics are available for No-Reference Point Cloud Quality Assessment (NR-PCQA). ResSCNN [ 12 ] proposes using voxelization to merge nearby points into a single voxel. Chetouani et al. [ 20 ] take a different approach by extracting hand-crafted features at the patch level and using a traditional CNN model for quality regression. PQA-net [ 3 ] and IT-PCQA [ 21 ] employ multiview projection for feature extraction. Zhang et al. [ 13 ] adopt a unique strategy, utilizing statistical distributions to estimate quality-related parameters from the geometry and attribute color distributions. Fan et al. attempt to estimate the visual quality of point clouds by analyzing captured video sequences [ 22 ] . Liu et al. [ 12 ] utilize an end-to-end sparse CNN to predict quality. Yang et al. [ 21 ] further advance this field by transferring quality scores from natural images to point cloud rendering images using domain adaptation techniques. However, in CNNs, the gradient of one kernel point’s weight during backpropagation is only related to the spatially corresponding isolated point in the patch due to the element-wise multiplication between the convolution kernel and the patch. This limited point interaction hampers the thorough extraction of structural and textural features [ 23 ] . Current no-reference metrics often ignore the rotation invariance, shift, and scale of point clouds, leading to inconsistent objective ratings for the same sample under different scales or positions. This oversight contrasts with most PCQA subjective studies, where observers have the freedom to choose the viewing angle and point cloud position [ 12 , 24 , 25 , 26 ] . Ignoring these invariances can reduce the robustness of the metric. Recently, graphs have received significant attention for point cloud processing applications [ 27 , 28 , 29 ] . Much of this progress is attributed to the ability of graph-based representations to model high-dimensional visual data, such as 3D point clouds, by implicitly capturing local neighbor relationships and highlighting their importance. Graphs are highly flexible data structures, making them ideal for representing 3D point clouds in an interpretable manner. The connectivity between a vertex and its neighboring vertices provides valuable insights into the perceptual characteristics of a point cloud. The application of graph representations has demonstrated significant efficacy in addressing various computer vision tasks, including image segmentation [ 30 ] and classification [ 31 ] . El Hassouni et al. [ 32 ] use graph feature learning and random forest regression to evaluate the visual quality of colored meshes. The key contributions of this work are outlined as follows: 1. The paper proposes a new no-reference point cloud quality assessment (PCQA) method called the Perceptual Clustering Weighted Graph (PCW-Graph). 2. The method begins by segmenting distorted point clouds into regions through clustering based on perceptual features. The clusters serve as nodes in the graph, and the ties between nodes are represented by weights that reflect the similarity of the cluster centers. 3. A Graph Attention Fusion Network (GAF) is employed during model training to better represent features, understand connections between perceptual features, and allow for flexible weight adjustments. 4. We compare the proposed method with several full-reference and state-of-the-art No-Reference PCQA metrics. The structure of this paper is as follows: Section 2 describes the proposed method, Section 3 details the experimental results, and Section 4 concludes the paper. 2 Related work Based on the kind of data they use, existing PCQA methods are generally categorized into two types: projection-based and point cloud-based methods. 2.1 Point Cloud-Based Methods Point cloud-based methods utilize raw point cloud data directly, extracting geometric, color, and other features to evaluate quality. Mekuria and Tian et al. were among the first to introduce point-based quality assessment methods, namely PSNR_p2po [ 33 ] and PSNR_p2pl [ 3 ] , marking the early development of PCQA methodologies. Later, PSNR_yuv [ 34 ] was introduced to assess texture distortion in color point clouds. Alexiou et al. [ 35 ] used the angle between corresponding points to characterize the degradation of distorted point clouds. Meynet et al. [ 36 ] proposed the MSDM method, which relies on local curvature statistics to measure point cloud quality. Javaheri et al. [ 37 ] employed the generalized Hausdorff distance in their PCQA work. In addition to geometric features, color