Research map / Label-efficient and robust learning / Domain adaptation and generalization
Mean Teacher DETR with Masked Feature Alignment: A Robust Domain Adaptive Detection Transformer Framework
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 MTM, a two-stage unsupervised domain adaptive DETR framework that pretrains on source and CycleGAN target-like images with masked domain-query and token-wise feature alignment, then self-trains with mean teacher and Object Queries Knowledge Transfer.
Problem
Unsupervised domain adaptive object detection with DETR suffers from performance fluctuation and training stagnation in one-stage feature alignment, and from unreliable pretrained teachers and unstable gains in mean-teacher two-stage methods; target-like images are unexplored for DETR adaptation.
Approach
Pretrain Deformable DETR on labeled source and CycleGAN target-like images with masked (randomly dropped) domain-query and token-wise feature discriminators to avoid stagnation, then self-train with mean teacher where student object queries are enhanced by attending to teacher queries (OQKT).
Key results
- Weather adaptation (Cityscapes to Foggy Cityscapes): MTM reaches 48.9 mAP vs. 46.8 for O2Net, a 2.1 mAP gain.
- Scene adaptation (Cityscapes to BDD100K): MTM reaches 37.3 mAP vs. 32.6 for MTTrans, a 4.7 mAP gain.
- Sim10K to Cityscapes: MTM reaches 58.1 car AP vs. 57.9 for MTTrans, a marginal 0.2 gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MTM instead of the strongest listed baseline | Deformable DETR · Cityscapes → Foggy Cityscapes | mAP: 46.8 -> 48.9 | Helped |
| Use MTM instead of the strongest listed baseline | Deformable DETR · Cityscapes → BDD100K | mAP: 32.6 -> 37.3 | Helped |
| Use MTM instead of MTTrans | Deformable DETR · Sim10k → Cityscapes | AP on car: 57.9 -> 58.1 | Helped |
| Add labeled target-like images to source-only pretraining (ablation) | Deformable DETR · Cityscapes → Foggy Cityscapes (pretraining) | mAP: 28.5 -> 42.3 | Helped |
| Replace non-masked MTWFA with masked MTWFA (ablation) | Target-like pretraining · Cityscapes → Foggy Cityscapes (pretraining) | mAP: 42.7 -> 43.6 | Helped |
| Replace non-masked MDQFA with masked MDQFA (ablation) | Target-like pretraining · Cityscapes → Foggy Cityscapes (pretraining) | mAP: 44.9 -> 45.7 | Helped |
| Replace both non-masked alignment methods with masked methods (ablation) | Target-like pretraining · Cityscapes → Foggy Cityscapes (pretraining) | mAP: 45.3 -> 47.2 | Helped |
| Set θMask=0.40 and η=1/2 (ablation) | MTM pretraining · Cityscapes → Foggy Cityscapes (pretraining) | mAP: 45.6 -> 47.2 | Helped |
| Use mean teacher self-training (ablation) | Deformable DETR · Cityscapes → Foggy Cityscapes (self-training) | mAP: 28.5 -> 35.8 | Helped |
| Add OQKT to mean teacher (ablation) | Deformable DETR self-training · Cityscapes → Foggy Cityscapes (self-training) | mAP: 35.8 -> 38.3 | Helped |
| Add OQKT to mean teacher with MTM pretraining (ablation) | MTM self-training · Cityscapes → Foggy Cityscapes | mAP: 47.7 -> 47.9 | Helped |
| Add masked MDQFA (ablation) | MTM self-training with mean teacher and OQKT · Cityscapes → Foggy Cityscapes | mAP: 47.9 -> 48.0 | Helped |
Surprising findings
- Pretraining with CycleGAN target-like images alone raises weather-adaptation mAP from 28.5 to 42.3, a 13.8 point gain.
- Masked feature alignment is reported to also help the pretraining stage beyond non-masked alignment, but the Sim10K gain over MTTrans is only 0.2 AP.
Limitations
- Gains over the strongest baseline are small in the synthetic-to-real scenario (0.2 AP on car).
- Ablations are only reported on the weather adaptation scenario.
- Theoretical analysis is loosely applied; the bound's assumptions are not validated empirically.
Experimental setup
Datasets: Cityscapes Foggy Cityscapes BDD100K Sim10K ImageNet
Benchmarks: Cityscapes to Foggy Cityscapes Cityscapes to BDD100K Sim10K to Cityscapes
Metrics: mAP@0.5 AP on car
Baselines: Deformable DETR (source only) SFA O2Net MTTrans DA-DETR
Builds on: Deformable DETR CycleGAN Mean Teacher MTTrans SFA
Closest papers by meaning
- Cross-Domain Adaptive Teacher for Object Detection (CVPR 2022 · similarity 0.91)
- Exploring Object Relation in Mean Teacher for Cross-Domain Detection (CVPR 2019 · similarity 0.89)
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (AAAI 2025 · similarity 0.89)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.88)
- Differential Alignment for Domain Adaptive Object Detection (AAAI 2025 · similarity 0.88)
- D3T: Distinctive Dual-Domain Teacher Zigzagging Across RGB-Thermal Gap for Domain-Adaptive Object Detection (CVPR 2024 · 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.