Information Exploration of Projected Views for Point Cloud Quality Measurement
Haohui Li, Wei Gao
Abstract
High-quality point clouds play a critical role in providing a satisfying visual experience in immersive media. Consequently, numerous methods have been developed for point cloud quality measurement (PCQM), which can be categorized into point- and projection-based approaches. Recently, existing projection-based methods have shown promising performance, and almost all of them treat the projection images of the input point cloud independently. However, these methods rarely consider the relationship among projected images and attempt to dig out the shared information among views. In order to leverage the shared cross-view information, we propose ViewPCQM, a new projection-based no-reference PCQM (NR-PCQM) metric, incorporating a novel multiview pooling (MVP) module. The MVP module comprises two submodules: cross-view spatial fusion (CVSF) and set pooling (SP). CVSF is designed to extract the cross-view information, while SP is proposed to pool the feature maps of the input multiview images into a single enhanced one. Then, the enhanced feature map can be transformed into a more descriptive feature vector. Our experiments show that the proposed metric achieves remarkable predicting performance by regressing with the feature vector obtained. We perform experiments on Shanghai Jiao Tong University Subjective Point Cloud Quality Assessment (SJTU-PCQA), Waterloo Point Cloud (WPC), and Waterloo Point Cloud 2.0 (WPC2.0) databases, and our method outperforms the existing PCQM methods, even including several full reference PCQM (FR-PCQM) methods. Furthermore, we verify the effectiveness of the MVP module in ablation experiments. Compared to the baseline model without the MVP module, ViewPCQM achieves significant performance gains, which suggests that cross-view information among views is worth exploring for the PCQM task. The source code will be available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://openi.pcl.ac.cn/OpenPointCloud</uri>.