Research map / Detection and segmentation / Open-set and few-shot detection
Unbiased Faster R-CNN for Single-source Domain Generalized 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 Unbiased Faster R-CNN (UFR), a causal-inspired single-source domain generalized object detector combining frequency/SAM-based augmentation with attention-invariance and prototype constraints on Faster R-CNN.
Problem
Single-source domain generalization for object detection under weather-induced distribution shift, where biased training data leads detectors to learn non-causal, poorly generalizable features.
Approach
Formulates SDG for detection with a structural causal model identifying data bias, attention bias and prototype bias from scene and object-attribute confounders, and mitigates them via Global-Local Transformation augmentation, a causal attention invariance loss, and causal prototype learning with explicit and implicit constraints.
Key results
- Achieves 40.8% mAP on Night-Clear, outperforming CLIP-Gap by 3.9% and vanilla Faster R-CNN by 5.0%.
- Achieves 39.6% mAP on Daytime-Foggy, 6.1% above SDGOD and 1.1% above CLIP-Gap.
- Achieves 58.6% mAP on Daytime-Clear, 2.4% above vanilla Faster R-CNN, and 19.2% on Night-Rainy, 0.5% above CLIP-Gap.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UFR instead of CLIP-Gap | Object detection on Night-Clear · Night-Clear | mAP@0.5: 36.9 -> 40.8 | Helped |
| Use UFR instead of CLIP-Gap | Object detection on Dusk-Rainy · Dusk-Rainy | mAP@0.5: 32.3 -> 33.2 | Helped |
| Use UFR instead of CLIP-Gap | Object detection on Night-Rainy · Night-Rainy | mAP@0.5: 18.7 -> 19.2 | Helped |
| Use UFR instead of CLIP-Gap | Object detection on Daytime-Foggy · Daytime-Foggy | mAP@0.5: 38.5 -> 39.6 | Helped |
| Add Global-Local Transformation (ablation) | Faster R-CNN · Daytime-Clear | mAP@0.5: 56.2 -> 60.4 | Helped |
| Add Global-Local Transformation (ablation) | Faster R-CNN · Night-Clear | mAP@0.5: 35.8 -> 38.6 | Helped |
| Add Global-Local Transformation (ablation) | Faster R-CNN · Daytime-Foggy | mAP@0.5: 33.5 -> 36.1 | Helped |
| Add Causal Attention Learning, Causal Prototype Learning, and their losses to GLT (ablation) | Faster R-CNN · Daytime-Clear | mAP@0.5: 60.4 -> 58.6 | Hurt |
| Add Causal Attention Learning, Causal Prototype Learning, and their losses to GLT (ablation) | Faster R-CNN · Night-Clear | mAP@0.5: 38.6 -> 40.8 | Helped |
| Add Causal Attention Learning, Causal Prototype Learning, and their losses to GLT (ablation) | Faster R-CNN · Daytime-Foggy | mAP@0.5: 36.1 -> 39.6 | Helped |
Surprising findings
- Adding the causal losses lowers source-domain (Daytime-Clear) performance while improving unseen-domain results, which the authors attribute to suppressing domain-specific knowledge.
- Attention invariance loss contributes more to generalization than the prototype losses, suggesting scene confounders matter more than object-level ones.
Limitations
- Gains on some scenes are modest (e.g., Night-Rainy 0.5% over CLIP-Gap, Daytime-Clear source-domain performance drops with causal losses in ablation)
- Evaluated on a single dataset with one source domain and five weather conditions
Experimental setup
Datasets: Daytime-Clear source domain (dataset from SDGOD)
Benchmarks: Daytime-Clear Night-Clear Dusk-Rainy Night-Rainy Daytime-Foggy
Metrics: mAP@0.5
Baselines: Faster R-CNN SW IBN-Net IterNorm ISW SDGOD CLIP-Gap
Builds on: Faster R-CNN Segment Anything Model CLIP frequency-domain augmentation
Closest papers by meaning
- Towards Single-Source Domain Generalized Object Detection via Causal Visual Prompts (NeurIPS 2025 · similarity 0.91)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.89)
- Generalized Diffusion Detector: Mining Robust Features from Diffusion Models for Domain-Generalized Detection (CVPR 2025 · similarity 0.88)
- G-NAS: Generalizable Neural Architecture Search for Single Domain Generalization Object Detection (AAAI 2024 · similarity 0.88)
- Object-Aware Domain Generalization for Object Detection (AAAI 2024 · similarity 0.88)
- Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery (CVPR 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.