4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks
Christopher Choy, JunYoung Gwak, Silvio Savarese
TL;DR
The paper tackles the inefficiency of processing 3D videos with frame-wise 2D/3D methods by introducing 4D spatio-temporal convolutional neural networks built on generalized sparse convolutions over sparse tensors. It presents the Minkowski Engine, an open-source library implementing flexible kernel shapes (including the hybrid kernel) and core neural-network operations, enabling scalable 4D networks. A 7D Trilateral Space-Time-Chroma CRF is proposed to enforce spatio-temporal-color consistency, trained end-to-end via variational inference. Empirical results on ScanNet, S3DIS, RueMonge/VarCity, and synthetic Synthia 4D datasets show that 4D networks can outperform 3D methods, are robust to noise, and can offer runtime advantages in certain settings. These contributions advance high-dimensional perception for robotics and AR/VR by making 4D, sparse-convolutional pipelines practical and accessible.
Abstract
In many robotics and VR/AR applications, 3D-videos are readily-available sources of input (a continuous sequence of depth images, or LIDAR scans). However, those 3D-videos are processed frame-by-frame either through 2D convnets or 3D perception algorithms. In this work, we propose 4-dimensional convolutional neural networks for spatio-temporal perception that can directly process such 3D-videos using high-dimensional convolutions. For this, we adopt sparse tensors and propose the generalized sparse convolution that encompasses all discrete convolutions. To implement the generalized sparse convolution, we create an open-source auto-differentiation library for sparse tensors that provides extensive functions for high-dimensional convolutional neural networks. We create 4D spatio-temporal convolutional neural networks using the library and validate them on various 3D semantic segmentation benchmarks and proposed 4D datasets for 3D-video perception. To overcome challenges in the 4D space, we propose the hybrid kernel, a special case of the generalized sparse convolution, and the trilateral-stationary conditional random field that enforces spatio-temporal consistency in the 7D space-time-chroma space. Experimentally, we show that convolutional neural networks with only generalized 3D sparse convolutions can outperform 2D or 2D-3D hybrid methods by a large margin. Also, we show that on 3D-videos, 4D spatio-temporal convolutional neural networks are robust to noise, outperform 3D convolutional neural networks and are faster than the 3D counterpart in some cases.