Contrastive Instance Association for 4D Panoptic Segmentation Using Sequences of 3D LiDAR Scans

Abstract

Scene understanding is critical for autonomous navigation in dynamic environments. Perception tasks in this domain like segmentation and tracking are usually tackled individually. In this letter, we address the problem of 4D panoptic segmentation using LiDAR scans, which requires to assign to each 3D point in a temporal sequence of scans a semantic class and for each object a temporally consistent instance ID. We propose a novel approach that builds on top of an arbitrary single-scan panoptic segmentation network and extends it to the temporal domain by associating instances across time. We propose a contrastive aggregation network that leverages the point-wise features from the panoptic network. It generates an embedding space in which encodings of the same instance at different timesteps lie close together and far from encodings belonging to other instances. The training is inspired by contrastive learning techniques for self-supervised metric learning. Our association module combines appearance and motion cues to associate instances across scans, allowing us to perform temporal perception. We evaluate our proposed method on the SemanticKITTI benchmark and achieve state-of-the-art results even without relying on pose information.