Search papers, labs, and topics across Lattice.
This paper introduces ChronoTrack, a novel 3D Single Object Tracking (3D-SOT) framework designed to overcome limitations in temporal feature consistency and memory overhead that plague existing memory-based approaches. ChronoTrack employs a compact set of learnable memory tokens coupled with a temporal consistency loss and a memory cycle consistency loss to ensure robust long-term target modeling. Experiments demonstrate state-of-the-art performance on multiple 3D-SOT benchmarks, achieving real-time speeds of 42 FPS.
By learning to maintain a consistent memory of target features across time, ChronoTrack achieves state-of-the-art 3D single object tracking at real-time speeds.
3D Single Object Tracking (3D-SOT) aims to localize a target object across a sequence of LiDAR point clouds, given its 3D bounding box in the first frame. Recent methods have adopted a memory-based approach to utilize previously observed features of the target object, but remain limited to only a few recent frames. This work reveals that their temporal capacity is fundamentally constrained to short-term context due to severe temporal feature inconsistency and excessive memory overhead. To this end, we propose a robust long-term 3D-SOT framework, ChronoTrack, which preserves the temporal feature consistency while efficiently aggregating the diverse target features via long-term memory. Based on a compact set of learnable memory tokens, ChronoTrack leverages long-term information through two complementary objectives: a temporal consistency loss and a memory cycle consistency loss. The former enforces feature alignment across frames, alleviating temporal drift and improving the reliability of proposed long-term memory. In parallel, the latter encourages each token to encode diverse and discriminative target representations observed throughout the sequence via memory-point-memory cyclic walks. As a result, ChronoTrack achieves new state-of-the-art performance on multiple 3D-SOT benchmarks, demonstrating its effectiveness in long-term target modeling with compact memory while running at real-time speed of 42 FPS on a single RTX 4090 GPU. The code is available at https://github.com/ujaejoon/ChronoTrack