No-reference point cloud quality assessment based on multi-projection and hierarchical pyramid network

Within the domains of 3D vision and computer graphics, point clouds play a pivotal role as a paramount technology in diverse applications. However, practical processes involved in their acquisition, storage, and transmission inevitably introduce reduction in quality of point clouds, which lead to point cloud quality assessment a pressing issue. Unfortunately, current research on no-reference methods for evaluating point cloud quality tend to disregard multi-scale feature analysis, resulting in suboptimal performance. Therefore, we present a novel no-reference approach that leverages multi-projection and a hierarchical pyramid network for evaluating the quality of point clouds. Specifically, the point cloud is initially projected onto multiple 2D images. Subsequently, a feature pyramid network (FPN) is employed to dissect the point cloud into distinct layers, enabling the extraction of features across various scales. These features are then fed into a quality regression module for obtaining predicted scores at multiple levels. Finally, the comprehensive point cloud quality score is derived by computing a weighted average of the predicted scores across various levels, thereby encapsulating information across multiple scales. The proposed method is implemented on SJTU-PCQA and WPC databases, aiming at assessing the quality of distorted point clouds. Our experiments demonstrate that the proposed method surpasses traditional no-reference point cloud quality assessment methods in terms of both accuracy and generalization.