Research map / Detection and segmentation / Tracking
Hybrid-SORT: Weak Cues Matter for Online 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.
Hybrid-SORT adds training-free weak cues (tracklet confidence, height-modulated IoU, and robust observation-centric velocity direction) to SORT-style trackers to improve association under occlusion and clustering.
Problem
Multi-object tracking association fails when occlusion and clustering make strong cues (spatial IoU and appearance) ambiguous due to high object overlap.
Approach
Incorporate weak cues (detection/tracklet confidence state, height state, and velocity direction) to compensate for strong cues: Kalman/linear-predicted tracklet confidence as an extra cost, height-modulated IoU (HIoU multiplied with IoU), and multi-interval, four-corner velocity direction, all in a plug-and-play, training-free SORT framework.
Key results
- Hybrid-SORT achieves 62.2 HOTA on DanceTrack test, and Hybrid-SORT-ReID reaches 65.7 HOTA, a state-of-the-art result among heuristic trackers there.
- On MOT20 test Hybrid-SORT-ReID reaches 63.9 HOTA, and on MOT17 test Hybrid-SORT-ReID reaches 64.0 HOTA, both highest among heuristic matchers reported.
- TCM improves DeepSORT by 4.9 HOTA on DanceTrack-val and HMIoU improves SORT by 1.6 HOTA on DanceTrack, showing gains across five trackers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Hybrid-SORT cues and modeling to OC-SORT | OC-SORT · DanceTrack test | HOTA: 54.6 -> 62.2 | Helped |
| Add Hybrid-SORT cues and modeling to OC-SORT | OC-SORT · MOT20 test | HOTA: 62.1 -> 62.5 | Helped |
| Add Hybrid-SORT cues and modeling to OC-SORT | OC-SORT · MOT17 test | HOTA: 63.2 -> 63.6 | Helped |
| Add ROCM (ablation) | Hybrid-SORT without added modules · DanceTrack-val | HOTA: 53.1 -> 53.7 | Helped |
| Add TCM (ablation) | Hybrid-SORT with ROCM · DanceTrack-val | HOTA: 53.7 -> 57.7 | Helped |
| Add HMIoU (ablation) | Hybrid-SORT with ROCM and TCM · DanceTrack-val | HOTA: 57.7 -> 59.3 | Helped |
| Add ReID (ablation) | Hybrid-SORT with ROCM, TCM, and HMIoU · DanceTrack-val | HOTA: 59.3 -> 63.0 | Helped |
| Use Kalman confidence modeling in the first association stage (ablation) | TCM · DanceTrack-val | HOTA: 53.7 -> 56.6 | Helped |
| Use Linear Prediction confidence modeling in the first association stage instead of Kalman (ablation) | TCM · DanceTrack-val | HOTA: 53.7 -> 52.6 | Hurt |
| Use Kalman confidence modeling in the second association stage (ablation) | TCM · DanceTrack-val | HOTA: 56.6 -> 57.3 | Helped |
| Use Linear Prediction confidence modeling in the second association stage (ablation) | TCM · DanceTrack-val | HOTA: 56.6 -> 57.7 | Helped |
| Replace height state with width state in the modulated IoU (ablation) | Hybrid-SORT · DanceTrack-val | HOTA: 59.3 -> 52.6 | Hurt |
Surprising findings
- Width-modulated IoU hurts association (HOTA drops from 57.7 to 52.6 on DanceTrack-val), while height-modulated IoU helps.
- Linear prediction of confidence decreases HOTA for high-confidence detections (-1.1), whereas Kalman filter is better there; the two are complementary.
Limitations
- Gains on MOT17/MOT20 are limited, attributed to small datasets and mostly linear motion.
- ReID integration reduces speed to near real-time, and efficient ReID integration is out of scope.
Experimental setup
Datasets: DanceTrack MOT17 MOT20
Benchmarks: DanceTrack MOT17 MOT20
Metrics: HOTA IDF1 MOTA FPS
Baselines: OC-SORT ByteTrack SORT DeepSORT MOTDT GHOST MOTRv2 SUSHI MotionTrack FineTrack BoT-SORT
Builds on: SORT OC-SORT ByteTrack BoT-SORT DeepSORT
Closest papers by meaning
- Focusing on Tracks for Online Multi-Object Tracking (CVPR 2025 · similarity 0.89)
- More Than Meets the Eye: Enhancing Multi-Object Tracking Even with Prolonged Occlusions (ICML 2025 · similarity 0.87)
- Multiple Object Tracking as ID Prediction (CVPR 2025 · similarity 0.87)
- Towards Generalizable Multi-Object Tracking (CVPR 2024 · similarity 0.87)
- Sampling-Resilient Multi-Object Tracking (AAAI 2024 · similarity 0.87)
- OVT-B: A New Large-Scale Benchmark for Open-Vocabulary Multi-Object Tracking (NeurIPS 2024 · similarity 0.87)
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.