An Investigation on Graph Convolution Networks for Color Enhancements of V-PCC

Dynamic Point Cloud (DPC) technology represents realistic 3D scenes in motion and has a wide range of applications. The Moving Picture Experts Group (MPEG) has developed Video-based Point Cloud Compression (V-PCC), achieving remarkable DPC compression performance. However, V-PCC introduces issues such as point reduction and coordinate distortion in the decoded DPC during the compression. In our previous work [14], we proposed an interpolation architecture to address the issue of point reduction by performing interpolation and coordinate adjustment on the decoded DPC. However, the attributes of the interpolated points were not addressed. In this paper, as a continuation of our previous work, we introduce a graph convolution network with an attention module incorporating geometric features to perform the color adjustment. Experimental results demonstrate that the interpolated DPC exhibits significant improvements in color quality, as evidenced by both objective and subjective evaluations. The highest PSNR gains compared to the interpolation architecture were 0.26 dB for the "redandblack"sequence and 0.17 dB for the "soldier"sequence.