Research map / Detection and segmentation / Open-set and few-shot detection
Differential Alignment for 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 PDFA (instance alignment weighted by teacher-student prediction discrepancy) and UFOA (uncertainty-weighted foreground/background image alignment) on top of an adaptive teacher-student Faster R-CNN for domain adaptive object detection.
Problem
Domain adaptive object detection via adversarial feature alignment treats all regions and instances equally, ignoring that different instances and foreground vs background regions carry different amounts of domain-specific information.
Approach
Weight instance-level adversarial alignment by teacher-student prediction discrepancy (PDFA), and split image-level features into foreground (pseudo-label/GT box mask) and background with a balancing factor gamma (UFOA), so both regions are aligned with differential attention rather than uniformly.
Key results
- Achieves 57.3% AP50 on Cityscapes to Foggy Cityscapes, vs 52.1% for REACT (second best).
- Achieves 45.8% AP50 on Cityscapes to BDD100K-daytime, 5.6 points above HT and 10.0 above REACT.
- Achieves 69.7% car AP50 on Sim10k to Cityscapes, 4.2 points above the previous best (REACT 58.6... PT 55.1 listed; text claims 4.2 margin over prior SOTA).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use differential alignment with PDFA and UFOA | Faster R-CNN · Cityscapes → Foggy Cityscapes (0.02) | AP50: 52.1 -> 57.3 | Helped |
| Use differential alignment with PDFA and UFOA | Faster R-CNN · Cityscapes → BDD100k-daytime | AP50: 45.8% AP50, outperforming ... HT by 5.6% AP50 | Helped |
| Use differential alignment with PDFA and UFOA | Faster R-CNN · Sim10k → Cityscapes (car) | car AP50: 69.7% AP50 ... by a margin of 4.2% | Helped |
| Add MIC strong augmentation (ablation) | Baseline teacher-student model with image- and instance-level alignment · Cityscapes → Foggy Cityscapes | AP50: 50.0 -> 53.9 | Helped |
| Replace equal instance alignment with PDFA (ablation) | Strong-augmentation baseline with UFOA · Cityscapes → Foggy Cityscapes | AP50: 55.2 -> 57.3; 2.1% | Helped |
| Replace equal image alignment with UFOA (ablation) | Strong-augmentation baseline with PDFA · Cityscapes → Foggy Cityscapes | AP50: 56.2 -> 57.3; 1.1% | Helped |
| Increase UFOA foreground weight from γ=0 to γ=0.5 (ablation) | UFOA · Cityscapes → Foggy Cityscapes | AP50: 55.1 -> 55.9 | Helped |
| Increase UFOA foreground weight from γ=0.5 to γ=0.8 (ablation) | UFOA · Cityscapes → Foggy Cityscapes | AP50: 55.9 -> 57.3 | Helped |
| Set UFOA foreground weight to γ=1.0, removing background alignment (ablation) | UFOA · Cityscapes → Foggy Cityscapes | AP50: 57.3 -> 55.6 | Hurt |
Surprising findings
- Setting gamma=1.0 (foreground-only alignment as in O^2net) lowers AP50 below gamma=0.8, indicating background alignment helps.
Limitations
- Only evaluated with Faster R-CNN; applicability to Deformable DETR untested.
- Pseudo-label-based foreground mask can be inaccurate for target domain.
- PDFA replaces teacher proposals with student proposals for consistency, an approximation.
Experimental setup
Datasets: Cityscapes Foggy Cityscapes Sim10k BDD100K-daytime
Benchmarks: Cityscapes to Foggy Cityscapes Cityscapes to BDD100K-daytime Sim10k to Cityscapes
Metrics: AP50
Baselines: FCOS EPM SIGMA CSDA HT Def DETR SFA O^2net MRT Faster R-CNN DA-Faster SADA
Builds on: Adaptive Teacher Faster R-CNN O^2net MIC
Closest papers by meaning
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (AAAI 2025 · similarity 0.93)
- DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.90)
- DSD-DA: Distillation-based Source Debiasing for Domain Adaptive Object Detection (ICML 2024 · similarity 0.90)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.89)
- Mean Teacher DETR with Masked Feature Alignment: A Robust Domain Adaptive Detection Transformer Framework (AAAI 2024 · similarity 0.88)
- Triple Feature Disentanglement for One-Stage Adaptive Object Detection (AAAI 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.