Research map / Detection and segmentation / CNN features
SelfMOTR: Revisiting MOTR with Self-Generating Detection Priors
Fabian Gülhan, Emil Mededovic, Yuli Wu, Johannes Stegmaier
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 self-generated proposals (decoupled proposal pass) (ablation) | MOTR · DanceTrack val | HOTA: 51.2 -> 58.2; AssA 38.4 -> 47.4 | Improved |
| add anchor proposals from detection-pretrained MOTR (ablation) | MOTR · DanceTrack val | HOTA: 51.2 -> 58.0 | Improved |
| add detection pretraining (ablation) | MOTR · DanceTrack val | HOTA: 51.2 -> 51.5; MOTA 74.4 -> 73.8 | No clear effect |
| add query pretraining (frozen detect queries) (ablation) | MOTR · DanceTrack val | AssA: 38.4 -> 41.1; IDF1 49.1 -> 51.9 | Improved |
| add detection distillation from frozen teacher (ablation) | MOTR · DanceTrack val | DetA: 68.8 -> 72.0; MOTA 74.4 -> 80.1; AssA 38.4 -> 37.9 | Mixed |
| use self proposal instead of anchor proposals from pretrained MOTR (ablation) | SelfMOTR · DanceTrack val | HOTA: 58.0 -> 58.2 | No clear effect |
| use 4D anchor proposals instead of learnable 4D anchors (ablation) | MOTR · DanceTrack val | HOTA: 55.8 -> 58.0 | Improved |
| add HSV augmentation and CrowdHuman joint training (ablation) | SelfMOTR · DanceTrack val | HOTA: 58.2 -> 64.3 | Improved |
| add denoising stage (ablation) | SelfMOTR · DanceTrack val | HOTA: 64.3 -> 64.4 | No clear effect |
| share decoder weights between proposal and tracking passes (ablation) | SelfMOTR · DanceTrack test | HOTA: 66.2 -> 69.2; AssA 54.5 -> 59.3 | Improved |
| lower proposal confidence threshold from 0.50 to 0.05 (ablation) | SelfMOTR · DanceTrack val | HOTA: 57.7 -> 58.2 | Improved |
| add track queries to MOTR detection (ablation) | MOTR · DanceTrack val | AP: 66.3 -> 60.0 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 5 on most shared results.
- MOTRv3 — MOTR: End-to-End Multiple-Object Tracking with TRansformer (ECCV 2022) · better on 4 of 4 shared results
- MeMOTR — MeMOTR: Long-Term Memory-Augmented Transformer for Multi-Object Tracking (ICCV 2023) · better on 2 of 3 shared results
- MOTRv2 — MOTRv2: Bootstrapping End-to-End Multi-Object Tracking by Pretrained Object Detectors (CVPR 2023) · better on 1 of 2 shared results
- ByteTrack — ByteTrack: Multi-Object Tracking by Associating Every Detection Box (ECCV 2022) · better on 2 of 2 shared results
- QDTrack — QDTrack: Quasi-Dense Similarity Learning for Appearance-Only Multiple Object Tracking (TPAMI 2023) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TrackFormer: Multi-Object Tracking With Transformers (CVPR 2022 · similarity 0.89)
- LA-MOTR: End-to-End Multi-Object Tracking by Learnable Association (ICCV 2025 · similarity 0.89)
- Compact Transformer Tracker with Correlative Masked Modeling (AAAI 2023 · similarity 0.89)
- Multiple Object Tracking as ID Prediction (CVPR 2025 · similarity 0.88)
- DeconfuseTrack: Dealing with Confusion for Multi-Object Tracking (CVPR 2024 · similarity 0.87)
- TGFormer: Transformer with Track Query Group for Multi-Object Tracking (AAAI 2025 · 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.
