Research map / Detection and segmentation / Open-set and few-shot detection
Object-Aware Domain Generalization for 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 OA-DG, a single-domain generalization method for object detection combining OA-Mix (multi-level transformations with saliency-based object-aware mixing) and OA-Loss (JS consistency plus object-aware supervised contrastive loss covering foreground and background instances).
Problem
Single-domain generalization for object detection: standard S-DG augmentations (designed for classification) can damage object locations and semantic features, and existing contrastive DG methods ignore background instances, hurting robustness to out-of-distribution corruptions and weather shifts.
Approach
Object-aware data augmentation that applies multi-level color/spatial transformations per region and mixes original and augmented images with weights set by per-object saliency scores, combined with a contrastive loss that treats background instances as positives only for their own augmented version rather than as one class.
Key results
- On Cityscapes-C, OA-DG reaches 43.4 clean mAP and 21.8 mPC with Faster R-CNN, versus 42.2 mAP and 15.4 mPC for the baseline.
- On the Diverse Weather Dataset, OA-DG achieves 31.8 mPC over OOD weathers, the best among compared S-DGOD methods (CDSD 28.7, SRCD 29.6).
- With YOLOv3, OA-DG improves mAP from 34.6 to 37.2 and mPC from 12.4 to 15.6 on Cityscapes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply OA-DG rather than the baseline | Faster R-CNN, ResNet-101 · DWD | mPC: 26.2 -> 31.8 | Helped |
| Apply OA-DG rather than the baseline | Faster R-CNN · Cityscapes | mPC: 12.4 -> 15.6 | Helped |
| Add single-level transformation without mixing (ablation) | OA-Mix · Cityscapes-C | mPC: 15.4 -> 16.8 | Helped |
| Add multi-level instead of single-level transformation (ablation) | OA-Mix · Cityscapes-C | mPC: 16.8 -> 17.5 | Helped |
| Add multi-level transformation without mixing (ablation) | OA-Mix · Cityscapes-C | mAP: 42.2 -> 40.5 | Hurt |
| Add standard mixing to multi-level transformation (ablation) | OA-Mix · Cityscapes-C | mPC: 17.5 -> 19.7 | Helped |
| Replace standard mixing with object-aware mixing (ablation) | OA-Mix · Cityscapes-C | mAP: 41.6 -> 42.7 | Helped |
| Replace standard mixing with object-aware mixing (ablation) | OA-Mix · Cityscapes-C | mPC: 19.7 -> 20.8 | Helped |
| Add consistency loss (ablation) | OA-Mix · Cityscapes-C | mPC: 20.8 -> 21.2 | Helped |
| Add foreground class-aware contrastive loss after consistency loss (ablation) | OA-Mix · Cityscapes-C | mAP: 43.1 -> 43.4 | Helped |
| Add foreground class-aware contrastive loss after consistency loss (ablation) | OA-Mix · Cityscapes-C | mPC: 21.2 -> 21.3 | Helped |
| Use object-aware rather than class-aware contrastive relations for foreground and background instances (ablation) | OA-Loss · Cityscapes-C | mPC: 21.1 -> 21.8 | Helped |
Surprising findings
- Several existing augmentations (Cutout, photometric distortion, AutoAug-det) gave limited or no gains on corrupted domains, and class-aware contrastive losses did not improve corrupted-domain performance.
- Multi-level transformations alone reduced clean mAP below the baseline (39.9 and 40.5 vs 42.2), and standard image-level mixing still lowered clean performance.
Limitations
- Mixing weights and saliency-based object-aware strategy add design complexity over standard augmentations.
- Gains on some DWD weather conditions (e.g., night-rainy) are small or negative relative to CDSD/SRCD.
- Evaluated mainly on urban driving benchmarks (Cityscapes-C and DWD) with Faster R-CNN and YOLOv3.
Experimental setup
Datasets: Cityscapes Cityscapes-C Diverse Weather Dataset BDD-100k FoggyCityscapes Adverse-Weather
Benchmarks: Cityscapes-C Diverse Weather Dataset
Metrics: mAP mPC
Baselines: AugMix Cutout Photo' AutoAug-det Stylized augmentation SupCon FSCE SW IBN-Net IterNorm ISW SHADE
Builds on: AugMix supervised contrastive learning FSCE Faster R-CNN PRIME
Closest papers by meaning
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.90)
- G-NAS: Generalizable Neural Architecture Search for Single Domain Generalization Object Detection (AAAI 2024 · similarity 0.89)
- Generalized Diffusion Detector: Mining Robust Features from Diffusion Models for Domain-Generalized Detection (CVPR 2025 · similarity 0.89)
- Towards Single-Source Domain Generalized Object Detection via Causal Visual Prompts (NeurIPS 2025 · similarity 0.89)
- Unbiased Faster R-CNN for Single-source Domain Generalized Object Detection (CVPR 2024 · similarity 0.88)
- Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection (NeurIPS 2024 · 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.