Research map / Detection and segmentation / Tracking
Delving into the Trajectory Long-tail Distribution for Muti-object Tracking
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.
Identifies a pedestrian trajectory long-tail distribution in MOT datasets and proposes SVA/DVA camera-view data augmentation plus a Group Softmax Re-ID loss, applied to FairMOT and CSTrack.
Problem
Multi-object tracking datasets have a long-tail distribution of trajectory lengths across pedestrian identities, so Re-ID learning is biased toward head classes (long trajectories) and under-learns tail classes (short trajectories).
Approach
Camera-view-aware augmentation: for stationary cameras, backtrack or Kalman-predict tail-class trajectories to extend them; for moving cameras, use SAM segmentation, Navier-Stokes inpainting and Stable Diffusion to change backgrounds while keeping pedestrians; plus a Group Softmax that partitions identities by sample count and applies softmax/cross-entropy per group.
Key results
- On MOT15, CSTrack+Ours improves HOTA from 42.9 to 46.4 and MOTA from 49.8 to 55.0.
- On MOT20 test, FairMOT trained only on MOT20 with the method reaches 65.9 MOTA and 70.3 IDF1, exceeding FairMOT trained on mixed data (61.8 MOTA, 67.3 IDF1).
- On MOT17 validation ablation, SVA+DVA+GS raises MOTA from 67.8 to 69.3 and IDF1 from 72.3 to 73.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SVA (ablation) | FairMOT · MOT17 validation set | MOTA: 67.8 -> 68.3 | Helped |
| Add SVA (ablation) | FairMOT · MOT17 validation set | IDF1: 72.3 -> 72.8 | Helped |
| Add DVA (ablation) | FairMOT · MOT17 validation set | MOTA: 67.8 -> 68.8 | Helped |
| Add DVA (ablation) | FairMOT · MOT17 validation set | IDF1: 72.3 -> 72.7 | Helped |
| Add SVA and DVA (ablation) | FairMOT · MOT17 validation set | MOTA: 67.8 -> 69.0 | Helped |
| Add SVA and DVA (ablation) | FairMOT · MOT17 validation set | IDF1: 72.3 -> 73.0 | Helped |
| Add GS to SVA and DVA (ablation) | FairMOT · MOT17 validation set | MOTA: 69.0 -> 69.3 | Helped |
| Add GS to SVA and DVA (ablation) | FairMOT · MOT17 validation set | IDF1: 73.0 -> 73.4 | Helped |
| Set DVA image selection threshold Ts to 0.9 (ablation) | FairMOT · MOT17 validation set | IDF1: 73.4 at 0.9; 72.5 at 0.7 | Mixed |
| Add SVA, DVA, and GS | FairMOT · MOT17 test set | HOTA: 57.1 -> 57.3 | Helped |
| Add SVA, DVA, and GS | FairMOT · MOT17 test set | IDF1: 70.2 -> 70.1 | Hurt |
Surprising findings
- Logit Adjustment and some re-balancing methods boost MOTA but reduce IDF1 or tail-class performance, and the Focal Loss and Triplet Loss variants hurt IDF1.
- On MOT16 and MOT17 test sets, gains are negligible (0 to 0.5 HOTA), with some metrics (e.g., FairMOT MOTA on MOT17) slightly decreasing.
Limitations
- Gains on MOT16 and MOT17 test sets are marginal compared with MOT15 and MOT20.
- Augmentation depends on hand-set thresholds (class threshold Tj, visibility Tv, selection threshold Ts) tuned per sequence.
- Evaluated only on two Joint Detection and Tracking trackers (FairMOT, CSTrack).
Experimental setup
Datasets: MOT15 MOT16 MOT17 MOT20 ETH CityPerson CalTech CUHK-SYSU PRW CrowdHuman
Benchmarks: MOT15 MOT16 MOT17 MOT20
Metrics: MOTA IDF1 HOTA MT ML IDS
Baselines: FairMOT CSTrack Focal Loss Triplet Loss CB Loss Logit Adjustment RelationTrack MTrack
Builds on: FairMOT CSTrack Segment Anything Model Stable Diffusion Logit adjustment
Closest papers by meaning
- DiffusionTrack: Diffusion Model for Multi-Object Tracking (AAAI 2024 · similarity 0.85)
- GigaTraj: Predicting Long-term Trajectories of Hundreds of Pedestrians in Gigapixel Complex Scenes (CVPR 2024 · similarity 0.85)
- Focusing on Tracks for Online Multi-Object Tracking (CVPR 2025 · similarity 0.84)
- DiffMOT: A Real-time Diffusion-based Multiple Object Tracker with Non-linear Prediction (CVPR 2024 · similarity 0.84)
- UCMCTrack: Multi-Object Tracking with Uniform Camera Motion Compensation (AAAI 2024 · similarity 0.83)
- Multi-Scene Generalized Trajectory Global Graph Solver with Composite Nodes for Multiple Object Tracking (AAAI 2024 · similarity 0.83)
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.