No-Reference Point Cloud Quality Assessment via Contextual Point-Wise Deep Learning Network

For the processing of point clouds, an accurate assessment of the quality is essential. However, point cloud quality assessment has proven to be a difficult issue, especially when the pristine point clouds are unavailable. Most existing no-reference point cloud quality assessment methods adopt projection-based routes, which inevitably suffer from occlusion and misalignment, resulting in loss of information. Alternatively, this paper proposes a novel no-reference point cloud quality assessment method via a contextual point-wise deep learning network (CPW-Net). Compared with projection-based methods, it reduces information loss by learning features directly from point coordinates and attributes. In particular, CPW-Net utilizes an Offset Attention Feature Encoder (OAFE) module to extract local and contextual features. Experiment results demonstrate that the proposed method overwhelms most publicly available no-reference metrics on SJTU dataset and gains compatible performance in comparison with most full-reference methods.