Restoration with Base Point Prediction for Deep Point Cloud Geometry Compression

In recent years, compression coding of point cloud geometry using deep learning has been studied. In deep learning-based methods, a large reduction in the amount of code results in degradations that look like cracks and holes. In this paper, a novel restoration method from such visually noticeable degradation is proposed. Although many completion and up-sampling methods have been studied in the past, they are not suitable as a solution to this problem, because the decoder side does not know the location of the degradation and it does not how to fill in. Therefore, in the proposed method, the base points for up-sampling are predicted according to the distortion map. The distortion map is generated at the encoder side and the parameters of the process for predicting and restoring points are estimated based on the distortion map and are encoded into the stream. At the decoder side, the restoration process for degradation is carried out according to these parameters. Evaluation experiments show that points lost due to degradation can be restored and subjective quality can be improved.