Enhancing Point Cloud Quality Assessment with Grouped Convolutions: A Streamlined Approach Inspired by COPP-Net

In the dynamic landscape of 3D vision applications, the Point Cloud Quality Assessment (PCQA) has become a critical focus. This paper presents an enhanced version of COPP-Net. COPP-Net strategically divides a point cloud into patches, leveraging a Point Cloud Pre-processing Module to normalize spatial coordinates and employ Farthest Point Sampling (FPS) and K-Nearest Neighbor algorithms for efficient patch creation. The subsequent Patch Feature Generation Module utilizes local texture and 3D structure feature generation Adaptive R-Sampling KNN PointNet++ Network networks (ARKPt and ARKPs) based on the ARKP architecture. Notably, the Adaptive R-Sampling KNN PointNet++ Network (ARKP) network undergoes improvements, including grouped convolutions and block reduction, resulting in a remarkable 50% reduction in trainable parameters and enhanced computational efficiency. The Point Cloud Quality Regression Module predicts the overall point cloud quality score from patch features, employing a regression head with linear layers, batch normalization, and leaky ReLU layers. The Correlation Analysis Network (CORA) network further refines the assessment by estimating correlations between patch and overall point cloud quality, introducing correlation labels for improved accuracy. Experiments conducted on diverse datasets, including WPC, WPC2.0, and LS-PCQA, showcase the efficacy of the improved COPP-Net. Impressively, the introduced improvements result in a 20% decrease in one-epoch time for ARKP and a 10% decrease for CORA, while maintaining consistent model accuracy across all tested datasets.