Adaptive Learning-Based Bilateral Filtering for Voxel Recoloring in Point Cloud Compression

In point cloud compression, voxels need to be recolored while preserving the maximum quality of experience (QoE) possible. Current point cloud recoloring techniques usually assign the average color intensity of all points within a voxel to its centroid. However, this approach neglects the spatial distribution of points, which is crucial for preserving visual fidelity. In our paper, we propose a learning based bilateral filtering technique on point cloud data. The geometry and photometric features of the local voxel region are computed to predict optimal kernel sizes in the bilateral filters to achieve adaptive performance. Getting motivated from the guided filter, a learning-based network is implemented to learn the optimal kernel sizes by training the variance and the other attributes-features of the points. Simulation results demonstrate modest gains, and the scheme holds significant potential for future optimization.