Research map / Detection and segmentation / Open-set and few-shot detection
PhysAug: A Physical-guided and Frequency-based Data Augmentation for Single-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 PhysAug, a frequency-domain data augmentation derived from an atmospheric-optics perturbation model (global non-uniform illumination plus particle-induced local occlusion) to improve single-domain generalized object detection.
Problem
Single-domain generalized object detection (S-DGOD): training a detector on one source domain so it generalizes to unseen target domains such as adverse weather and corruptions, where existing augmentations ignore real-world imaging formation.
Approach
Derives a universal physical perturbation model from atmospheric optics and implements it as Fourier-domain augmentation: random Gaussian low-frequency filters simulate global illumination, and random sinusoidal planar waves simulate local particle occlusions, with no change to architecture or loss.
Key results
- On DWD, PhysAug reaches 37.5 mPC versus 30.2 for the baseline (+7.3), and the best results in Daytime Sunny (60.2 mAP), Night Sunny (44.9), Dusk Rainy (41.2) and Daytime Foggy (40.8).
- On Cityscapes-C, PhysAug improves mPC from 15.4 (baseline) to 22.6 (+7.2) and beats OA-DG (21.8 mPC), while keeping clean mAP at 42.6 versus 42.2 for the baseline.
- Ablation on Cityscapes-C shows physics-based PhysAug (22.6 mPC) outperforms the non-physical variants NPM1 (22.2) and NPM2 (19.4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply PhysAug instead of the baseline augmentation | Faster R-CNN ResNet101 · DWD adverse weather | mPC: +7.3 mPC | Helped |
| apply PhysAug instead of the baseline augmentation | Faster R-CNN ResNet50 FPN · Cityscapes-C | mPC: 7.2 mPC | Helped |
| add non-uniform illumination to local occlusion (ablation) | PhysAug · DWD | mPC: 35.2 -> 37.5 | Helped |
| add local occlusion to non-uniform illumination (ablation) | PhysAug · DWD | mPC: 36.5 -> 37.5 | Helped |
| add non-uniform illumination to local occlusion (ablation) | PhysAug · Cityscapes-C | mPC: 20.2 -> 22.6 | Helped |
| add local occlusion to non-uniform illumination (ablation) | PhysAug · Cityscapes-C | mPC: 20.5 -> 22.6 | Helped |
| use physical-based modeling rather than NPM1 (ablation) | PhysAug on Faster R-CNN ResNet50 FPN · Cityscapes-C | mPC: 22.2 -> 22.6 | Helped |
| use physical-based modeling rather than MixUp-style NPM2 (ablation) | PhysAug on Faster R-CNN ResNet50 FPN · Cityscapes-C | mPC: 19.4 -> 22.6 | Helped |
Surprising findings
- On Cityscapes-C, the illumination-only and occlusion-only variants reach about 20.2-20.5 mPC, while the combined PhysAug reaches 22.6 mPC, but the combination does not improve clean mAP over occlusion-only (42.6 vs 43.0 for occlusion-only).
- On DWD, the occlusion-only variant has the highest mAP (60.6) but lower mPC (35.2) than the full method (37.5).
Limitations
- Relies on hand-chosen distributions for filter, frequency, phase and attenuation parameters (e.g., uniform (-512,512), d in (0,10)) without a principled selection procedure.
- Evaluated only on DWD and Cityscapes-C with a Faster R-CNN detector; generalization to other detectors or datasets is not shown.
Experimental setup
Datasets: Diverse Weather Dataset (DWD) Cityscapes Cityscapes-C
Benchmarks: Diverse Weather Dataset (DWD) Cityscapes-C
Metrics: mAP mPC
Baselines: IBN-Net SW IterNorm ISW SHADE CDSD SRCD DivAlign OA-DG UFR Cutout PhotoDistort
Builds on: Retinex theory atmospheric scattering model MixUp CDSD OA-DG
Closest papers by meaning
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.87)
- Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery (CVPR 2025 · similarity 0.87)
- Towards Single-Source Domain Generalized Object Detection via Causal Visual Prompts (NeurIPS 2025 · similarity 0.87)
- Object-Aware Domain Generalization for Object Detection (AAAI 2024 · similarity 0.86)
- Unbiased Faster R-CNN for Single-source Domain Generalized Object Detection (CVPR 2024 · similarity 0.86)
- Generalized Diffusion Detector: Mining Robust Features from Diffusion Models for Domain-Generalized Detection (CVPR 2025 · similarity 0.86)
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.