Problem
Research questions and friction points this paper is trying to address.
End-to-End Multi-Object Tracking
Spatial Priors
Identity Association
Spatially Implausible Errors
MOTIP2: Spatial Priors for End-to-End Multi-Object Tracking
🤖 AI Summary
This study addresses the identity mismatch problem in end-to-end multi-object tracking caused by the absence of spatial priors. Building upon the DEIM detection transformer architecture, this work proposes explicit spatial prior strategies across three levels: data, loss, and representation. Specifically, it introduces Spatial ID Switches, Spatial ID Loss, and Spatial Anchor techniques to optimize trajectory association and attention mechanisms. These components effectively correct long-range identity switch errors without requiring additional annotations while preserving fully end-to-end inference. Experimental results demonstrate that the proposed method achieves state-of-the-art performance on benchmarks such as DanceTrack, attaining a HOTA score of 73.4. Furthermore, a lightweight variant of the model accelerates inference speed by more than threefold, offering an efficient yet highly accurate solution for real-time tracking applications.
📝 Abstract
End-to-end multi-object trackers have narrowed the gap with classical tracking-by-detection on association-difficult benchmarks. Yet they still make spatially implausible errors no classical tracker would, such as assigning one identity to objects on opposite sides of the frame. A model could learn to avoid them, but tracking annotations are scarce, so we encode spatial priors explicitly instead, while keeping inference fully end-to-end with no post-hoc association.
We propose three spatial priors, at the data, loss, and representation stages. Spatial ID Switches bias trajectory permutations toward spatially overlapping objects, reducing the mismatch between training and inference confusions. Spatial ID Loss scales each identity's penalty by its box distance, so a distant switch costs more than a nearby one. Spatial Anchor gives each track token its frame position, an explicit spatial cue for attention.
We instantiate the three priors in MOTIP2, a tracker adapted from MOTIP and built on the real-time DEIM detection transformer. Trained without extra data, its main model, MOTIP2-L, sets a new state of the art: 73.4 HOTA on DanceTrack, 76.0 on SportsMOT, and 71.1 IDF1 on PersonPath22. MOTIP2 is a family of models spanning the speed-accuracy trade-off: a lighter model, MOTIP2-S, matches the original MOTIP at over 3x the speed, and MOTIP2-X reaches 74.8 HOTA on DanceTrack.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Multi-Object Tracking
Spatial Priors
End-to-End
Spatial ID Loss
Detection Transformer
🔎 Similar Papers
No similar papers found.
