No-Reference Point Cloud Quality Assessment Based on Multi-Projection and Hierarchical Pyramid Network

Within the realms of 3D vision and computer graphics, point clouds play a pivotal role as a dominant technology in various applications. However, practical processes such as acquisition, storage, and transmission introduce diverse quality defects in point clouds, making point cloud quality assessment a pressing issue. Presently, research on no-reference methods for evaluating point cloud quality often overlooks multi-scale feature analysis, resulting in suboptimal performance. This study presents a novel approach for evaluating the quality of point clouds based on multi-projection and a hierarchical pyramid network without using any reference data. Specifically, the point cloud is initially projected onto multiple 2D images. Subsequently, a pyramid network is employed to decompose the point cloud into multiple layers, extracting features at various scales. These features are subsequently inputted into a quality regression module for obtaining predicted scores at diverse levels. Finally, a weighted average of predicted scores from different levels is calculated to derive a point cloud quality score that encompasses information from multiple scales. The proposed method is applied on two databases, namely SJTU-PCQA and WPC, to assess the quality of distorted point clouds. Our experiments demonstrate that the proposed method surpasses traditional NR-PCQA methods regarding accuracy and generalization.