Research map / Detection and segmentation / Open-set and few-shot detection
G-NAS: Generalizable Neural Architecture Search for Single Domain Generalization 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 G-NAS, a differentiable NAS applied to the class prediction head of Faster R-CNN, regularized by a G-loss (spectral-decoupling-style), to improve single-domain generalization object detection on urban-scene weather/time shifts.
Problem
Single domain generalization object detection (S-DGOD): training a detector on one source urban domain (Daytime-Sunny) and generalizing to multiple unseen target domains; differentiable NAS overfits due to easy-feature bias.
Approach
Applies differentiable NAS to the class prediction head and regularizes both detector and architecture parameters with an OoD-aware G-loss (difference of squared classification and regression output norms, inspired by spectral decoupling) so training does not collapse onto easy features.
Key results
- G-NAS reaches 33.5% average mAP over the four target domains, versus 29.6% for the prior SOTA SRCD (reported as up to 33.5% vs 29.6%).
- Removing NAS drops average mAP to 31.0% with G-loss; removing G-loss drops it to 31.1%, per the ablation.
- On Night-Rainy, G-NAS reaches 17.4% mAP, only 0.4% above SRCD, while the car AP improves by 6.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use G-NAS instead of SRCD | Faster R-CNN, ResNet-101 · Four target domains average | mAP: 29.6% -> 33.5% | Helped |
| use G-NAS instead of the SOTA baseline | Faster R-CNN, ResNet-101 · Daytime-Foggy | AP (person): +5.2% | Helped |
| use G-NAS instead of the SOTA baseline | Faster R-CNN, ResNet-101 · Dusk-Rainy | mAP: +5.6% | Helped |
| use G-NAS instead of SOTA baselines | Faster R-CNN, ResNet-101 · Night-Sunny | mAP: +8.3% | Helped |
| use G-NAS instead of SOTA baselines | Faster R-CNN, ResNet-101 · Night-Rainy | mAP: +0.4% | Helped |
| add NAS (ablation) | G-NAS · Four target domains average | mAP: +2.4% mAP with G-loss; +1.6% mAP without G-loss | Helped |
| add G-loss (ablation) | G-NAS · Four target domains average | mAP: +2.5% mAP and 1.7% mAP | Helped |
| set λ_g to 1 (ablation) | G-NAS · Four target domains average | mAP: 33.5% | Helped |
Surprising findings
- The paper reports that the NAS-based gains are modest (about 2.4 mAP) and the Night-Rainy gap to SOTA is marginal, which it attributes to limits of the base detector.
Limitations
- Theory relies on infinite-width NTK approximation and a single-bounding-box simplification, so the guarantee does not directly cover practical detectors.
- Convergence claim is marked TODO in the draft, and the search uses training loss rather than validation loss.
- Evaluated only on urban-scene weather/time shifts with a single source domain and one detector family.
Experimental setup
Datasets: BDD100K Cityscapes Foggy Cityscapes Adverse-Weather
Benchmarks: Daytime-Sunny Daytime-Foggy Dusk-Rainy Night-Sunny Night-Rainy
Metrics: mAP per-class AP (PASCAL VOC)
Baselines: Faster R-CNN IBN-Net Switchable Whitening IterNorm RobustNet (ISW) CDSD SHADE CLIPGap SRCD
Builds on: DARTS Faster R-CNN Spectral Decoupling NAS-OoD Neural Tangent Kernel
Closest papers by meaning
- Object-Aware Domain Generalization for Object Detection (AAAI 2024 · similarity 0.89)
- Unbiased Faster R-CNN for Single-source Domain Generalized Object Detection (CVPR 2024 · similarity 0.88)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.88)
- Towards Single-Source Domain Generalized Object Detection via Causal Visual Prompts (NeurIPS 2025 · similarity 0.88)
- Generalized Diffusion Detector: Mining Robust Features from Diffusion Models for Domain-Generalized Detection (CVPR 2025 · similarity 0.86)
- SoMA: Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning (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.