Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations

• DOI: 10.48550/arxiv.2403.13261
• Publication Date: 2024-03-19
• Authors: Kewei Wang; Yizheng Wu; Jun Cen; Zhiyu Pan; Xingyi Li; Zhe Wang; Zhiguo Cao; Guosheng Lin
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

The perception of motion behavior in a dynamic environment holds significant importance for autonomous driving systems, wherein class-agnostic prediction methods directly predict the entire point cloud. While most existing rely on fully-supervised learning, manual labeling cloud data is laborious and time-consuming. Therefore, several annotation-efficient have been proposed to address this challenge. Although effective, these weak annotations or additional multi-modal like images, potential benefits inherent sequence are still underexplored. To end, we explore feasibility self-supervised with only unlabeled LiDAR clouds. Initially, employ an optimal transport solver establish coarse correspondences between current future clouds as pseudo labels. Training models using such labels leads noticeable spatial temporal inconsistencies. mitigate issues, introduce three simple regularization losses, which facilitate training process effectively. Experimental results demonstrate superiority our approach over state-of-the-art methods.