Transform Decomposition Switching for Efficient Attribute Compression of 3D Point Clouds Using Neural Networks

An adaptive technique to switch between RAHT and Dyadic RAHT using 3D Sobel filter has been found to improve the compression in 3D point clouds by offering substantial cumulative compression gains. However, the drawback of this switching scheme is its need for tuned thresholds. To this end, we propose to use neural networks to resolve the threshold dependency issue so that the switching becomes truly adaptive. Two publicly available point cloud datasets were used to test the effectiveness of the proposed method. We achieved significant gains on MVUB and minor gains on 8iVFB dataset over all Dyadic approach.