Learnable Online Graph Representations for 3D Multi-Object Tracking

Autonomous systems that operate in dynamic environments require robust object tracking 3D as one of their key components. Most recent approaches for multi-object (MOT) from LIDAR use dynamics together with a set handcrafted features to match detections objects across multiple frames. However, manually designing such and heuristics is cumbersome often leads suboptimal performance. In this work, we instead strive towards unified learning based approach the MOT problem. We design graph structure jointly process detection track states an online manner. To end, employ Neural Message Passing network data association fully trainable. Our provides natural way initialization handling false positive detections, while significantly improving stability. demonstrate merit proposed nuScenes challenge 2021 state-of-the-art performance 65.6% AMOTA 58% fewer ID-switches, resulting best only submission overall second place.