A Spatiotemporal Correspondence Approach to Unsupervised LiDAR Segmentation with Traffic Applications

TL;DR

This paper presents an unsupervised method for LiDAR point cloud segmentation in traffic scenarios, leveraging spatiotemporal correspondences and pseudo-labels to learn discriminative features without supervision.

Contribution

It introduces a novel spatiotemporal correspondence approach combined with clustering and pseudo-label learning for unsupervised LiDAR segmentation.

Findings

Achieves competitive performance on Semantic-KITTI, SemanticPOSS, and FLORIDA datasets.

Demonstrates effective unsupervised learning of discriminative features.

Provides a unified framework for LiDAR point cloud representation learning.

Abstract

We address the problem of unsupervised semantic segmentation of outdoor LiDAR point clouds in diverse traffic scenarios. The key idea is to leverage the spatiotemporal nature of a dynamic point cloud sequence and introduce drastically stronger augmentation by establishing spatiotemporal correspondences across multiple frames. We dovetail clustering and pseudo-label learning in this work. Essentially, we alternate between clustering points into semantic groups and optimizing models using point-wise pseudo-spatiotemporal labels with a simple learning objective. Therefore, our method can learn discriminative features in an unsupervised learning fashion. We show promising segmentation performance on Semantic-KITTI, SemanticPOSS, and FLORIDA benchmark datasets covering scenarios in autonomous vehicle and intersection infrastructure, which is competitive when compared against many existing fully supervised learning methods. This general framework can lead to a unified representation learning approach for LiDAR point clouds incorporating domain knowledge.