Point Cloud Quality Assessment Based on Cross-Modal De-Redundancy

Xiaoying Ding, Bin Luo, Hanjiang Xiong, Dawei Jin, Jun Wan, Li Jiang

Abstract: 3D point cloud has been widely applied in multimedia field, such as virtual reality, smart city and etc. However, during the process of 3D point cloud acquisition, storage and transmission, lots of noises will be introduced, leading to degraded visual quality. The accurate quantification for the visual quality of the distorted 3D point cloud remains a big challenge. In this paper, we propose a novel no-reference point cloud quality assessment (NR-PCQA) metric based on cross-modal de-redundancy. First, the 2D projections of the 3D point cloud are generated to obtain 2D quality-aware features. Then, the 3D point cloud is decomposed into patches to obtain 3D quality-aware features and a patch correlation-aware (PCA) module is utilized to explore the correlation information between 3D patches. Later, a cross-modal de-redundancy (CDR) module is designed which first reduces the information redundancy between features and then performs attention computation to explore the information between different modalities. Finally, a regression framework is applied to help obtain the quality evaluation scores. To examine the performance of our proposed PCQA metric, we compare it with several SOTA PCQA metrics using two classic databases. The comparison results prove the superior performance of our approach.

External IDs:doi:10.1109/access.2024.3521093