Research map / Detection and segmentation / Tracking
DeconfuseTrack: Dealing with Confusion 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 Decomposed Data Association (DDA), a non-learning, plug-in association refinement with three disambiguation modules, plus Occlusion-aware NMS (ONMS), yielding the DeconfuseTrack tracker that improves ByteTrack on MOT17/MOT20.
Problem
Global data association in multi-object tracking retains confusion among detections, trajectories and associations (ID switches, wrong assignments), and standard NMS discards occluded detections that could aid association.
Approach
Decompose global Hungarian association into detection-, trajectory- and association-level disambiguation sub-problems (DDM, TDM, ADM) that use motion cues first and selectively add appearance cues within a confusion-reduction factor, combined with occlusion-aware NMS that keeps a second tier of occluded detections for association.
Key results
- On MOT17 test (private detection), DeconfuseTrack reaches 64.9 HOTA, 80.6 IDF1, 65.1 AssA, improving over ByteTrack's 63.1 HOTA, 77.3 IDF1.
- On MOT20 test, DeconfuseTrack achieves 63.3 HOTA and 77.6 IDF1 versus ByteTrack's 61.3 and 75.2.
- On MOT17 validation, DDA and ONMS together raise ByteTrack-based HOTA from 69.0 to 70.7 and IDF1 from 80.6 to 83.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DDA and ONMS to ByteTrack | ByteTrack · MOT17 test | HOTA: 63.1 -> 64.9 | Helped |
| Add DDA and ONMS to ByteTrack | ByteTrack · MOT17 test | IDF1: 77.3 -> 80.6 | Helped |
| Add DDA and ONMS to ByteTrack | ByteTrack · MOT20 test | HOTA: 61.3 -> 63.3 | Helped |
| Add DDA and ONMS to ByteTrack | ByteTrack · MOT20 test | IDF1: 75.2 -> 77.6 | Helped |
| Add DDM alone (ablation) | ByteTrack · MOT17 val | HOTA: 69.0 -> 69.1 | Helped |
| Add TDM alone (ablation) | ByteTrack · MOT17 val | HOTA: 69.0 -> 69.7 | Helped |
| Add ADM alone (ablation) | ByteTrack · MOT17 val | HOTA: 69.0 -> 69.5 | Helped |
| Add DDM and TDM (ablation) | ByteTrack · MOT17 val | HOTA: 69.0 -> 69.9 | Helped |
| Add DDM, TDM, and ADM (ablation) | ByteTrack · MOT17 val | MOTA: 79.4 -> 80.4 | Helped |
| Add ONMS alone (ablation) | ByteTrack · MOT17 val | MOTA: 79.4 -> 80.0 | Helped |
| Add DDA alone (ablation) | ByteTrack · MOT17 val | IDF1: 80.6 -> 82.1 | Helped |
| Add DDA and ONMS (ablation) | ByteTrack · MOT17 val | IDF1: 80.6 -> 83.7 | Helped |
Surprising findings
- DDM alone yields only a small gain, but combined with TDM it gives a notable improvement because freed high-score detections are re-associated.
- DDA shows little or slightly negative IDF1 effect on DeepSORT, which uses appearance cues already.
Limitations
- Sensitive to the confusion reduction factor kappa, which must be tuned per dataset motion characteristics.
- Gains on appearance-based trackers such as DeepSORT are limited.
Experimental setup
Datasets: MOT17 MOT20
Benchmarks: MOT17 MOT20
Metrics: HOTA IDF1 MOTA AssA DetA
Baselines: ByteTrack OC-SORT StrongSORT++ QuoVadis SAT UTM MOTRv2 FairMOT
Builds on: ByteTrack SORT DeepSORT FastReID
Closest papers by meaning
- Multiple Object Tracking as ID Prediction (CVPR 2025 · similarity 0.88)
- Focusing on Tracks for Online Multi-Object Tracking (CVPR 2025 · similarity 0.87)
- ADA-Track: End-to-End Multi-Camera 3D Multi-Object Tracking with Alternating Detection and Association (CVPR 2024 · similarity 0.87)
- More Than Meets the Eye: Enhancing Multi-Object Tracking Even with Prolonged Occlusions (ICML 2025 · similarity 0.87)
- DiffusionTrack: Diffusion Model for Multi-Object Tracking (AAAI 2024 · similarity 0.86)
- Multi-Scene Generalized Trajectory Global Graph Solver with Composite Nodes for Multiple Object Tracking (AAAI 2024 · similarity 0.86)
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.