Reproduce “FutrTrack: A Camera-LiDAR Fusion Transformer for 3D Multiple Object Tracking” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Martha Teiko Teye, Ori Maoz, Matthias Rottmann
FutrTrack presents a modular transformer tracker for 3D multi-object tracking that combines LiDAR geometry with camera-derived bird’s-eye-view features. On nuScenes, camera-LiDAR fusion raises validation aMOTA from 0.736 to 0.759 over a matched LiDAR-only baseline, while reducing identity switches from 392 to 340; temporal smoothing further improves trajectory stability, though performance remains dependent on detector quality and generally trails the strongest geometry-based trackers.
Key insights
- Camera information improves tracking association beyond its role in object detection, with the clearest gains appearing for bicycles, motorcycles, and other objects that may be sparsely represented by LiDAR.
- Temporal smoothing primarily reduces localization jitter and fragmentation; expanding the smoothing window improves aMOTA and reduces identity switches, but does not compensate for missed detections.
- LiDAR remains the more critical modality: removing it lowers aMOTA to 0.568, whereas removing an individual camera causes a smaller degradation, indicating complementary but unequal sensor contributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.