Research map / Detection and segmentation / Open-set and few-shot detection
Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment
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 diversifying a single source domain with visual corruptions (ImageNet-C plus Fourier) and aligning Faster R-CNN/FCOS detections between clean and augmented views via KL on class distributions and L2 on boxes, for single-source domain-generalized object detection.
Problem
Single-source domain generalization for object detection, where a detector trained on one labeled source domain (Pascal VOC or BDD daytime-sunny) must generalize to unseen target domains without target data, and remain well calibrated.
Approach
Combines a strong augmentation-based baseline (ImageNet-C and Fourier corruptions, excluding weather and constant amplitude) with a detector-agnostic consistency loss that aligns class distributions (KL divergence) and box regressions between original and augmented views, using shared RoI proposals for correspondence.
Key results
- On Pascal VOC to Clipart1k/Watercolor2k/Comic2k, the full method reaches 38.9/57.4/33.2 mAP versus 25.7/44.5/18.9 for Faster R-CNN (gains of about 13.2, 12.9, 14.3 points).
- On urban scene detection (Daytime Sunny source), the method averages best among compared methods, e.g. 42.5 on Night Clear and 24.1 on Night Rainy versus 38.9 and 15.7 for the baseline.
- Alignment lowers D-ECE calibration error versus diversification alone (e.g. Clipart 10.7 vs 14.5 for diversification-only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add diversification to the Faster R-CNN baseline (ablation) | Faster R-CNN · Pascal VOC → Clipart1k | mAP@0.5 (%): 25.7 -> 34.2 | Helped |
| Add diversification to the Faster R-CNN baseline (ablation) | Faster R-CNN · Pascal VOC → Watercolor2k | mAP@0.5 (%): 44.5 -> 53.0 | Helped |
| Add diversification to the Faster R-CNN baseline (ablation) | Faster R-CNN · Pascal VOC → Comic2k | mAP@0.5 (%): 18.9 -> 24.2 | Helped |
| Add classification and localization alignment to diversification (ablation) | Faster R-CNN · Pascal VOC → Clipart1k | mAP@0.5 (%): 34.2 -> 38.9 | Helped |
| Add classification and localization alignment to diversification (ablation) | Faster R-CNN · Pascal VOC → Watercolor2k | mAP@0.5 (%): 53.0 -> 57.4 | Helped |
| Add classification and localization alignment to diversification (ablation) | Faster R-CNN · Pascal VOC → Comic2k | mAP@0.5 (%): 24.2 -> 33.2 | Helped |
| Use classification alignment only, without localization alignment (ablation) | Diversification (Faster R-CNN) · Pascal VOC → Clipart1k | mAP@0.5 (%): 38.9 -> 36.2 | Hurt |
| Use localization alignment only, without classification alignment (ablation) | Diversification (Faster R-CNN) · Pascal VOC → Clipart1k | mAP@0.5 (%): 38.9 -> 35.0 | Hurt |
| Use both ImageNet-C and Fourier-based corruptions rather than either alone (ablation) | Ours (Faster R-CNN) · Pascal VOC → Comic2k | mAP@0.5 (%): 30.6 -> 33.2 | Helped |
| Use both ImageNet-C and Fourier-based corruptions rather than Fourier-based corruptions alone (ablation) | Ours (Faster R-CNN) · Pascal VOC → Comic2k | mAP@0.5 (%): 29.1 -> 33.2 | Helped |
| Add diversification and classification/localization alignment | Faster R-CNN · Real to artistic shifts | D-ECE (%) | Helped |
Surprising findings
- Simple diversification with augmentations alone degrades calibration (higher D-ECE) even as it improves mAP, attributed to NLL-trained overconfidence on corrupted images.
- Diversification-only produces more confident false positives (e.g., spurious chair and bird detections), indicating augmentation alone is insufficient.
Limitations
- Evaluated only on VOC-to-artistic and a single urban driving benchmark with one source domain.
- Alignment requires two forward passes per image (clean and augmented), increasing training cost.
- Some hyperparameters (alpha, beta) and detailed sensitivity are not fully characterized in the main text.
Experimental setup
Datasets: Pascal VOC 2007 Pascal VOC 2012 Clipart1k Watercolor2k Comic2k Urban Scene Detection (BDD-100k, FoggyCityscapes, Adverse-Weather) ImageNet-C corruptions Fourier-based corruptions
Benchmarks: Pascal VOC to Clipart1k/Watercolor2k/Comic2k Urban Scene Detection (Daytime Sunny to Night Clear/Night Rainy/Dusk Rainy/Daytime Foggy)
Metrics: mAP@0.5 D-ECE AP per class
Baselines: Faster R-CNN FCOS NP SW IBN-Net IterNorm ISW Wu et al. (cyclic-disentangled self-distillation) Vidit et al. (text-prompt semantic augmentation) DA-Faster SWDA HTCN
Builds on: Faster R-CNN FCOS ImageNet-C Normalization Perturbation
Closest papers by meaning
- Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection (NeurIPS 2024 · similarity 0.90)
- Object-Aware Domain Generalization for Object Detection (AAAI 2024 · similarity 0.90)
- Towards Single-Source Domain Generalized Object Detection via Causal Visual Prompts (NeurIPS 2025 · similarity 0.89)
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (AAAI 2025 · similarity 0.89)
- DSD-DA: Distillation-based Source Debiasing for Domain Adaptive Object Detection (ICML 2024 · similarity 0.89)
- Unbiased Faster R-CNN for Single-source Domain Generalized Object Detection (CVPR 2024 · similarity 0.89)
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.