Graph-based Network for Dynamic Point Cloud Prediction

Abstract

Dynamic point clouds have enabled the rise of virtual reality applications. However, due to their voluminous size, point clouds require efficient compression methods. While a few articles have addressed the compression of dynamic point clouds by exploring temporal redundancies between sequential frames, very few have explored point cloud prediction as a tool for efficient compression. In this PhD thesis, we propose an end-to-end learning network to predict future frames in a point cloud sequence. To address the challenges present in point cloud processing, namely the lack of structure we propose a graph-based approach to learn topological information of point clouds as geometric features. Early results demonstrate that our method is able to make accurate predictions and can be applied in a compression algorithm.