Research map / Detection and segmentation / Tracking
UCMCTrack: Multi-Object Tracking with Uniform Camera Motion Compensation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UCMCTrack, a motion-only multi-object tracker that runs a Kalman filter on the ground plane with a Mapped Mahalanobis Distance replacing IoU, and applies one fixed camera-compensation setting per sequence instead of frame-by-frame CMC.
Problem
Multi-object tracking under significant camera motion, where image-plane motion models and IoU association fail and frame-by-frame camera motion compensation (CMC) adds substantial computational cost that hinders real-time tracking.
Approach
Model target motion with a constant-velocity Kalman filter on the ground plane (via homography from manually estimated camera parameters), map detection noise to a correlated ground-plane covariance, and associate with a normalized Mahalanobis distance (MMD) instead of IoU; camera motion is treated as process noise using a fixed set of compensation factors for the whole sequence.
Key results
- UCMCTrack+ reaches 65.8 HOTA and 81.1 IDF1 on MOT17 test, surpassing prior SOTA by 0.9 HOTA, 0.5 IDF1 and 0.7 AssA.
- On DanceTrack test, UCMCTrack+ improves over prior SOTA by 2.3 HOTA, 3.4 IDF1 and 5.5 AssA.
- UCMCTrack runs above 1000 FPS on a single CPU when given detections.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace IoU association with MMD without CMD (ablation) | UCMCTrack · MOT17 validation | HOTA: 68.43 -> 68.09 | Hurt |
| add CMD to MMD (ablation) | UCMCTrack · MOT17 validation | HOTA: 68.09 -> 70.44 | Helped |
| add process noise compensation (PNC) (ablation) | UCMCTrack · MOT17 validation | HOTA: 70.44 -> 71.96 | Helped |
| replace IoU association with MMD without CMD (ablation) | UCMCTrack · DanceTrack validation | HOTA: 47.93 -> 46.32 | Hurt |
| add CMD to MMD (ablation) | UCMCTrack · DanceTrack validation | HOTA: 46.32 -> 55.06 | Helped |
| add process noise compensation (PNC) (ablation) | UCMCTrack · DanceTrack validation | HOTA: 55.06 -> 60.42 | Helped |
| add CMC (ablation) | baseline tracker · MOT17 validation | HOTA: 68.43 -> 70.97 | Helped |
| add CMC (ablation) | UCMCTrack · MOT17 validation | HOTA: 71.96 -> 72.97 | Helped |
| add CMC (ablation) | baseline tracker · DanceTrack validation | HOTA: 47.93 -> 47.55 | Hurt |
| add CMC (ablation) | UCMCTrack · DanceTrack validation | HOTA: 60.42 -> 59.18 | Hurt |
| vary detector noise factor σm (ablation) | UCMCTrack · MOT17 validation | HOTA | Mixed |
Surprising findings
- Applying frame-by-frame CMC on DanceTrack lowered both baseline and UCMCTrack performance, which the authors attribute to minor detection offsets for overlapping targets.
- Replacing IoU with MMD alone decreased performance on MOT17 validation (IDF1 77.10 to 75.88), and gains only appear after adding correlated measurement distribution.
Limitations
- Assumes targets lie on a single ground plane.
- Camera parameters for MOT17, MOT20 and DanceTrack are manually estimated, and errors in X and Z extrinsics noticeably degrade performance.
- Uses motion cues only, without height information in MMD, and CMC hurt performance on KITTI and DanceTrack.
Experimental setup
Datasets: MOT17 MOT20 DanceTrack KITTI
Benchmarks: MOT17 MOT20 DanceTrack KITTI
Metrics: HOTA IDF1 MOTA AssA FP FN TA
Baselines: ByteTrack OC-SORT C-BIoU MotionTrack SparseTrack Quo Vadis FCG GHOST Bot-SORT StrongSORT Deep OCSORT QD-3DT
Builds on: ByteTrack OC-SORT BoT-SORT StrongSORT
Closest papers by meaning
- DiffusionTrack: Diffusion Model for Multi-Object Tracking (AAAI 2024 · similarity 0.87)
- Sampling-Resilient Multi-Object Tracking (AAAI 2024 · similarity 0.87)
- Focusing on Tracks for Online Multi-Object Tracking (CVPR 2025 · similarity 0.86)
- SMILEtrack: SiMIlarity LEarning for Occlusion-Aware Multiple Object Tracking (AAAI 2024 · similarity 0.86)
- Hybrid-SORT: Weak Cues Matter for Online Multi-Object Tracking (AAAI 2024 · similarity 0.85)
- Temporal Coherent Object Flow for Multi-Object Tracking (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.