Spatio-temporal Graph-RNN for Point Cloud Prediction

TL;DR

This paper introduces a novel end-to-end network that predicts future point cloud frames by learning topological and geometric features, utilizing Graph-RNN cells to model spatio-temporal dynamics effectively.

Contribution

The main novelty is an initial layer that captures topological information of point clouds, combined with Graph-RNN cells for dynamic modeling, improving prediction accuracy over geometry-agnostic methods.

Findings

Outperforms baseline methods on multiple datasets

Effectively captures spatio-temporal point cloud dynamics

Demonstrates robustness across different motion types

Abstract

In this paper, we propose an end-to-end learning network to predict future frames in a point cloud sequence. As main novelty, an initial layer learns topological information of point clouds as geometric features, to form representative spatio-temporal neighborhoods. This module is followed by multiple Graph-RNN cells. Each cell learns points dynamics (i.e., RNN states) by processing each point jointly with the spatio-temporal neighbouring points. We tested the network performance with a MINST dataset of moving digits, a synthetic human bodies motions and JPEG dynamic bodies datasets. Simulation results demonstrate that our method outperforms baseline ones that neglect geometry features information.