Research map / Label-efficient and robust learning / Domain adaptation and generalization
Towards Single-Source Domain Generalized Object Detection via Causal Visual Prompts
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 Cauvis, which pairs frozen DINOv2 with cross-attention visual prompts and a dual-branch (causal MLP + Fourier high-pass) adapter to reduce spurious-correlation reliance for single-source domain generalized object detection.
Problem
Single-source domain generalized object detection (SDGOD) models over-rely on spurious features such as color and spatial layout, which degrades performance on unseen target weather/domain conditions.
Approach
Cross-attention between trainable visual prompts and image features, interpreted as a back-door adjustment, combined with a dual-branch adapter in which one branch maps cross-attended prompt activations to causal features and an auxiliary Fourier branch applies high-pass masking to isolate domain-invariant structure, all on a frozen DINOv2 backbone.
Key results
- Cauvis reaches 60.7% average mAP on SDGOD with frozen DINOv2-L, versus 57.5% for frozen DINOv2 alone and 54.3% for DINO w/o prompts.
- On Cityscapes-C, Cauvis improves mPC over FR+DINOv2 from 29.3 to 35.6 (reported as 13.8% over prior SOTA).
- On BDD100K-C, FR+Cauvis raises mPC from 41.3 to 43.6 over FR+DINOv2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Cauvis rather than using FR + DINOv2 | Faster R-CNN on Cityscapes-C level-5 · Cityscapes-C level-5 | mPC: 29.3 -> 35.6 | Helped |
| add Cauvis rather than using FR + DINOv2 | Faster R-CNN on BDD100K-C level-5 · BDD100K-C level-5 | mPC: 41.3 -> 43.6 (+2.3) | Helped |
| add the dual-branch adapter (ablation) | Cauvis · SDGOD Avg. | mAP: 59.8 -> 60.7 | Helped |
| add cross-attention prompts (ablation) | Cauvis · SDGOD Avg. | mAP: 57.4 -> 60.7 | Helped |
| add visual prompts (ablation) | Cauvis · SDGOD Avg. | mAP: 54.3 -> 60.7 | Helped |
| use DINOv2 as the backbone (ablation) | Cauvis · SDGOD Avg. | mAP: 34.2 -> 60.7 | Helped |
| remove the Fourier high-pass mask (ablation) | Cauvis auxiliary branch · SDGOD Avg. | mAP: 60.7 -> 60.5 | Hurt |
| remove the Fourier transform (ablation) | Cauvis auxiliary branch · SDGOD Avg. | mAP: 60.7 -> 59.8 | Hurt |
| replace the causal branch with multi-head, SE Block, Gate Unit, or 3×3 Conv (ablation) | Cauvis · SDGOD Avg. | mAP: 60.7 -> 59.4; 58.9; 59.0; 58.6 | Hurt |
| increase prompt length from around 100 to sequence-aligned L_seq (ablation) | Cauvis · SDGOD prompt-length ablation | mAP: 59.7; 59.8 -> 59.9 (≤0.2 points) | Helped |
Surprising findings
- Plain DINOv2 backbone swap alone gives the largest gain, about 23 mAP over the DINO baseline on average, per the paper's ablation (34.2 to 57.5 average reported in ablation tables).
Limitations
- Spurious-correlation analysis is limited to a preliminary white truck/bus color bias experiment and does not cover other bias types.
- High training resource overhead: about 6 hours on eight RTX 4090 GPUs across multiple backbones.
- Theoretical back-door adjustment equivalence relies on strong assumptions (low-dimensional causal subspace, prompts spanning confounder space) that are approximate in practice.
Experimental setup
Datasets: SDGOD benchmark (Day Clear source) Cityscapes-C BDD100K-C Cityscapes
Benchmarks: SDGOD Cityscapes-C BDD100K-C CD-FSOD
Metrics: mAP@0.5 mPC
Baselines: UFR PDDOC ClipGap G-NAS SRCD CDSD OA-DG Rein SPT EVP VPT AdaptFormer
Builds on: DINO Rein VPT Faster R-CNN
Closest papers by meaning
- Unbiased Faster R-CNN for Single-source Domain Generalized Object Detection (CVPR 2024 · similarity 0.91)
- Style Evolving along Chain-of-Thought for Unknown-Domain Object Detection (CVPR 2025 · similarity 0.90)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.89)
- Object-Aware Domain Generalization for 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)
- Prompt-Driven Dynamic Object-Centric Learning for Single Domain Generalization (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.