Research map / Label-efficient and robust learning / Domain adaptation and generalization
Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection
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 DINO Teacher, which replaces the Mean Teacher labeller with a Faster R-CNN head on a frozen DINOv2 encoder trained on source only, and aligns student patch features to a frozen DINOv2 ViT-B to improve domain adaptive object detection.
Problem
Domain adaptive object detection, where Mean Teacher self-labelling couples label generation to the student being trained, producing brittle and poor target pseudo-labels, especially under large domain gaps and on small target datasets.
Approach
Decouple pseudo-label generation from the student by training an independent detector on a frozen large self-supervised DINOv2 encoder using only source data, and align student backbone patch features on source and target images to frozen DINOv2 features without using any labels, replacing the domain discriminator.
Key results
- DT achieves 47.8 mAP on Cityscapes to BDD100k, exceeding Harmonious Teacher (40.2) by 7.6 points.
- DT achieves 55.4 mAP on Cityscapes to Foggy Cityscapes, improving over REACT (53.1) by 2.3 points.
- On ACDC, DT-G improves over reimplemented Adaptive Teacher by +6.4 on fog, +6.9 on night, +1.3 on rain and +1.6 on snow.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DINO Teacher instead of Harmonious Teacher | DAOD from Cityscapes to BDD100k · Cityscapes → BDD100k | mAP@50: +7.6% | Helped |
| Use DINO Teacher instead of REACT | DAOD from Cityscapes to Foggy Cityscapes · Cityscapes → Foggy Cityscapes | mAP@50: +2.3% | Helped |
| Use DINO Teacher instead of Adaptive Teacher | DAOD on ACDC Fog, Night, Rain, and Snow · Cityscapes → ACDC splits | mAP@50: +6.4% on fog, +6.9% on night, +1.3% on rain and +1.6% on snow | Helped |
| Replace domain adversarial loss with patch similarity loss while retaining Mean Teacher labels (ablation) | Adaptive Teacher on BDD100k · Cityscapes → BDD100k | best mAP: 31.8 -> 35.3 | Helped |
| Replace Mean Teacher labels with DINO labeller while retaining domain adversarial alignment (ablation) | Adaptive Teacher on BDD100k · Cityscapes → BDD100k | best mAP: 31.8 -> 46.8 | Helped |
| Add DINO patch similarity loss to the full method (ablation) | DINO Teacher on BDD100k · Cityscapes → BDD100k | best mAP: 46.8 -> 47.8 | Helped |
| Increase DINO labeller backbone from ViT-B to ViT-G (ablation) | DINO Teacher on BDD100k · Cityscapes → BDD100k | student mAP: +1.8% | Helped |
| Align to DINOv2 ViT-B instead of ViT-S features (ablation) | DINO Teacher on BDD100k · Cityscapes → BDD100k | mAP: +0.4% | Helped |
| Use DINOv2 ViT-G labeller instead of source-only student model | Target-domain label generation · BDD100k | mAP@50: 29.1 -> 51.1 | Helped |
| Use DINO Teacher instead of target-only training | DAOD on BDD100k · BDD100k | mAP@50: 54.4 -> 47.8 | Hurt |
| Add source training to target training with ground-truth target labels | Detector trained on BDD100k · BDD100k | mAP@50: 54.4 -> 51.6 | Hurt |
Surprising findings
- The student trained with DINO labels can outperform its own labeller for ViT-B and ViT-L labellers, so labeller performance is not an upper bound.
- Training with source data plus full target GT labels reduces performance on BDD100k compared to target-only training.
Limitations
- Gains on small ACDC splits are limited by pseudo-label quality and overfitting to pseudo-labels on small target sets.
- DT still lags behind target-only training on the larger BDD100k dataset in the ACDC-style comparison, and remains below the ground-truth-label upper bound.
- Experiments use relatively small backbones (VGG16, ResNet-50) for the student, with large VFMs only used as labeller and alignment target.
Experimental setup
Datasets: Cityscapes BDD100k Foggy Cityscapes ACDC
Benchmarks: Cityscapes to BDD100k Cityscapes to Foggy Cityscapes Cityscapes to ACDC (fog, night, rain, snow)
Metrics: mAP@50
Baselines: Adaptive Teacher Harmonious Teacher (HT) REACT CMT DA-Faster SIGMA TDD PT NSA CAT MIC NLTE
Builds on: Adaptive Teacher Mean Teacher Faster R-CNN DINOv2 Unbiased Teacher
Closest papers by meaning
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (AAAI 2025 · similarity 0.91)
- DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.90)
- CAT: Exploiting Inter-Class Dynamics for Domain Adaptive Object Detection (CVPR 2024 · similarity 0.89)
- Differential Alignment for Domain Adaptive Object Detection (AAAI 2025 · similarity 0.89)
- SEEN-DA: SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection (CVPR 2025 · similarity 0.88)
- Towards Unsupervised Model Selection for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.88)
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.