Research map / Detection and segmentation / Tracking
Temporal Coherent Object Flow for Multi-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.
Proposes OFTrack, a section-based multi-object tracker that attaches a multi-frame object flow head to YOLOX to jointly estimate detection and multi-frame object motion, trained with bidirectional masked-trajectory pre-training and a curvature smoothing loss.
Problem
Frame-by-frame multi-object tracking relies on motion predictors and appearance cues that adapt poorly to long-term associations and complex, non-linear motion in crowded scenes.
Approach
Boosts optical flow to object flow: treats multiple consecutive frames as a section and jointly predicts object trajectories across the section using object-aware sampling, scale-aware correlation, and spatial-temporal attention, with a flow head sharing the detector backbone; adds bidirectional masked-track pre-training and a curvature smoothing constraint to stabilize long trajectory estimation.
Key results
- OFTrack reaches 60.9 HOTA on DanceTrack test, and 63.1 HOTA on MOT17 and 61.9 HOTA on MOT20 under the private protocol; with ReID these rise to 63.4, 64.1 and 63.4.
- On KITTI MOT, OFTrack reports HOTA 73.75 for cars and 47.97 for pedestrians.
- Ablation on MOT17 val-half: the full model with object perception, bidirectional pre-train and curvature constraint reaches 80.6 MOTA, 77.4 IDF1 and 66.7 HOTA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ReID module | OFTrack · DanceTrack test | HOTA: 60.9 -> 63.4 | Helped |
| add ReID module | OFTrack · MOT17 test | HOTA: 63.1 -> 64.1 | Helped |
| add ReID module | OFTrack · MOT20 test | HOTA: 61.9 -> 63.4 | Helped |
| use OFTrack instead of DiffMOT | DanceTrack test · DanceTrack test | HOTA: 62.3 -> 60.9 | Hurt |
| use OFTrack instead of MotionTrack | MOT17 test · MOT17 test | HOTA: 65.1 -> 63.1 | Hurt |
| use OFTrack instead of OC-SORT | MOT20 test · MOT20 test | HOTA: 62.4 -> 61.9 | Hurt |
| use OFTrack instead of Eager | KITTI Car tracking · KITTI MOT test, Car | HOTA: 74.39 -> 73.75 | Hurt |
| use OFTrack instead of MPNTrack | KITTI Person tracking · KITTI MOT test, Person | HOTA: 45.26 -> 47.97 | Helped |
| add target sampling (ablation) | OFTrack object perception · MOT17 val-half | HOTA: 51.9 -> 60.7 | Helped |
| add bidirectional pre-train with 6(8) masked tracks (ablation) | OFTrack · MOT17 val-half | HOTA: 63.3 -> 65.5 | Helped |
| add curvature constraint (ablation) | OFTrack with bidirectional pre-train · MOT17 val-half | HOTA: 65.5 -> 66.7 | Helped |
| increase motion-estimation iterations from 1 to 6 (ablation) | OFTrack · MOT17 val-half | HOTA: 59.9 -> 63.3 | Helped |
Surprising findings
- The paper reports that a rigid-object assumption causes tracking drift when target aspect ratio varies widely, limiting non-rigid deformation handling.
Limitations
- Treats targets as rigid objects, so it does not handle non-rigid deformation well.
- Possible tracking drift when target aspect ratio varies widely.
- Performance on very long sequences depends on section length; overly long sections degrade accuracy.
Experimental setup
Datasets: MOT17 MOT20 DanceTrack KITTI
Benchmarks: MOT17 MOT20 DanceTrack KITTI MOT
Metrics: HOTA MOTA IDF1 DetA AssA
Baselines: CenterTrack MOTR ByteTrack P3AFormer MeMOT GTR TransCenter OC-SORT MotionTrack DiffusionTrack DiffMOT TrackMPNN
Builds on: YOLOX RAFT-style optical flow correlation ByteTrack BoT-SORT
Closest papers by meaning
- Focusing on Tracks for Online Multi-Object Tracking (CVPR 2025 · similarity 0.87)
- DiffusionTrack: Diffusion Model for Multi-Object Tracking (AAAI 2024 · similarity 0.86)
- UCMCTrack: Multi-Object Tracking with Uniform Camera Motion Compensation (AAAI 2024 · similarity 0.85)
- FlowTrack: Revisiting Optical Flow for Long-Range Dense Tracking (CVPR 2024 · similarity 0.85)
- Multiple Object Tracking as ID Prediction (CVPR 2025 · similarity 0.85)
- ADA-Track: End-to-End Multi-Camera 3D Multi-Object Tracking with Alternating Detection and Association (CVPR 2024 · 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.