Research map / Detection and segmentation / Open-set and few-shot detection
DIOD: Self-Distillation Meets Object Discovery
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 DIOD, a self-distillation slot-attention method that refines noisy motion masks and adds static objects via teacher-student pseudo-labeling for unsupervised video object discovery.
Problem
Motion-guided object discovery relies on noisy, sparse optical-flow motion masks, which causes incomplete or merged objects, background noise, and missed static objects.
Approach
Places slot attention inside an EMA teacher-student distillation framework: teacher attention regions become one-to-many pseudo-labels for the student, motion masks are filtered by a learned foreground objectness map, and the foreground/background modeling is regularized over the whole map.
Key results
- On TRI-PD, DIOD reaches 66.1 fg-ARI versus 53.9 for BMOD; with DINOv2 (DIOD*) it reaches 69.7.
- On KITTI, DIOD reaches 73.5 fg-ARI versus 54.7 for BMOD and 64.4 for MoTok.
- On MOVi-E, DIOD* reaches 82.2 fg-ARI, above SLOV (80.8) and VideoSAUR (78.4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace BMOD with DIOD | Motion-guided object discovery · TRI-PD test | fg-ARI: 53.9 -> 66.1 | Helped |
| Replace BMOD with DIOD | Motion-guided object discovery · KITTI test | F1@50: 9.3 -> 18.0 | Helped |
| Use DINOv2 pre-training | DIOD · TRI-PD test | fg-ARI: 66.1 -> 69.7 | Helped |
| Use DINOv2 pre-training | DIOD · KITTI test | fg-ARI: 73.5 -> 72.3 | Hurt |
| Use one-to-one instead of one-to-many teacher-student attention connections (ablation) | DIOD · TRI-PD test | F1@50: 35.4 -> 30.2 | Hurt |
| Remove motion masks during teacher-student training (ablation) | DIOD · TRI-PD test | F1@50: 35.4 -> 17.8 | Hurt |
| Replace continuous teacher-student distillation with offline pseudo-labeling (ablation) | DIOD · TRI-PD test | F1@50: 35.4 -> 26.3 | Hurt |
| Add teacher-student distillation after burn-in (ablation) | DIOD · TRI-PD test, burn-in 200 epochs, α=0.2 | F1@50: 12.7 → 29.3 | Helped |
| Increase burn-in from 200 to 400 epochs at α=0.2 (ablation) | DIOD · TRI-PD test | F1@50: 15.6 → 23.6 | Hurt |
| Add teacher-student distillation after burn-in (ablation) | DIOD · TRI-PD test, burn-in 400 epochs, α=0.3 | F1@50: 16.9 → 35.4 | Helped |
| Increase burn-in from 200 to 400 epochs at α=0.3 (ablation) | DIOD · TRI-PD test | F1@50: 14.6 → 30.7; 16.9 → 35.4 | Helped |
| Increase burn-in from 200 to 400 epochs at α=0.4 (ablation) | DIOD · TRI-PD test | F1@50: 15.5 → 32.7; 18.8 → 33.7 | Mixed |
Surprising findings
- One-to-one teacher-student connection causes the model to drift from instance to semantic segmentation by merging nearby objects.
- Longer burn-in can hurt performance when noise suppression strength alpha is low, because weakly attenuated noise is recovered.
Limitations
- Regularization strength alpha is a fixed hyperparameter, not adapted per frame.
- Confidence score (mean activation) is a simple heuristic that does not account for object completeness.
- Motion masks are still required as a fixed supervision source; the method is not end-to-end.
Experimental setup
Datasets: ParallelDomain (TRI-PD) KITTI MOVi-E
Benchmarks: TRI-PD KITTI MOVi-E
Metrics: fg-ARI all-ARI F1@50
Baselines: Bao et al. BMOD MoTok SAVI SAVI++ PPMP STEVE VideoSAUR SLOV
Builds on: slot attention BMOD mean teacher (EMA) motion-guided object discovery
Closest papers by meaning
- Cross-Modal Distillation for 2D/3D Multi-Object Discovery from 2D Motion (CVPR 2025 · similarity 0.86)
- P2Seg: Pointly-supervised Segmentation via Mutual Distillation (ICLR 2024 · similarity 0.85)
- SeMoLi: What Moves Together Belongs Together (CVPR 2024 · similarity 0.85)
- Learning Segmentation from Point Trajectories (NeurIPS 2024 · similarity 0.85)
- The Devil is in the Object Boundary: Towards Annotation-free Instance Segmentation using Foundation Models (ICLR 2024 · similarity 0.85)
- ZeroFlow: Scalable Scene Flow via Distillation (ICLR 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.