Research map / Detection and segmentation / Open-set and few-shot detection
DARTH: Holistic Test-time Adaptation for Multiple Object Tracking
Mattia Segu, Bernt Schiele, Fisher Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DARTH test-time adaptation to QDTrack under sim-to-real shift | QDTrack · SHIFT -> BDD100K (average) | MOTA: -66.4 -> 8.3 | Improved |
| apply patch contrastive learning alone without detection consistency (ablation) | DARTH · SHIFT -> BDD100K (average) | DetA: 12.0 -> 9.4 | Hurt |
| add detection consistency alone (ablation) | DARTH · SHIFT -> BDD100K (average) | MOTA: -66.4 -> -37.6 | Improved |
| add EMA teacher updates to detection consistency (ablation) | DARTH · SHIFT -> BDD100K (average) | MOTA: -37.6 -> 6.1 | Improved |
| add patch contrastive learning to EMA + DC (ablation) | DARTH · SHIFT -> BDD100K (average) | MOTA: 6.1 -> 8.3 | Improved |
| generate student view with photometric augmentation from teacher view (ablation) | DARTH · SHIFT -> BDD100K (average) | MOTA: -5.8 -> 8.3 | Improved |
| add photometric augmentation to the teacher view (ablation) | DARTH · SHIFT -> BDD100K (average) | MOTA: -5.8 -> -6.8 | Hurt |
| use Tent (entropy minimization on BN) for adaptation | QDTrack · SHIFT -> BDD100K (average) | MOTA: -66.4 -> 0.0 | Hurt |
| use SFOD detector-only self-training for adaptation | QDTrack · MOT17 (+CH) -> BDD100K | MOTA: 28.3 -> 17.0 | Hurt |
| add DARTH to MOT17 (+CH) source model on BDD100K | QDTrack · MOT17 (+CH) -> BDD100K | MOTA: 28.3 -> 23.4 | Hurt |
| pretrain detector on CrowdHuman before training QDTrack on MOT17 (ablation) | QDTrack source model · MOT17 -> DanceTrack/BDD100K | DetA: up to +9.4 DetA and +6.8 AssA | Improved |
| DARTH versus SFOD on BDD100K -> MOT17 detection | DARTH · BDD100K -> MOT17 | DetA: 29.6 (SFOD) vs 29.4 | Mixed |
Closest papers by meaning
- Dual-level Adaptation for Multi-Object Tracking: Building Test-Time Calibration from Experience and Intuition (CVPR 2026 · similarity 0.89)
- Towards Real-Time Multi-Object Tracking (ECCV 2020 · similarity 0.87)
- Test-Time Adaptive Object Detection with Foundation Model (NeurIPS 2025 · similarity 0.85)
- Multiple Object Tracking as ID Prediction (CVPR 2025 · similarity 0.85)
- Hybrid-SORT: Weak Cues Matter for Online Multi-Object Tracking (AAAI 2024 · similarity 0.84)
- Towards Generalizable Multi-Object Tracking (CVPR 2024 · similarity 0.84)
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.
