Research map / Detection and segmentation / Open-set and few-shot detection
DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive 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 DA-Ada, a domain-aware adapter inserted into a frozen CLIP-based RegionCLIP detector that pairs a domain-invariant adapter with a domain-specific adapter fed by the residual the block discards, plus a visual-guided textual adapter, for domain adaptive object detection.
Problem
Domain adaptive object detection with frozen vision-language model visual encoders: domain-agnostic adapters learn domain-invariant features biased toward the source domain and discard target-discriminative domain-specific knowledge.
Approach
Attach a parallel domain-invariant adapter (adversarially aligned) to each frozen visual-encoder block and a domain-specific adapter that takes the block's discarded residual (input minus DIA output), fusing them via pixel-level attention; additionally, condition the textual classifier on visual embeddings through a visual-guided textual adapter.
Key results
- DA-Ada reaches 58.5% mAP on Cityscapes to Foggy Cityscapes, versus 55.9% for DA-Pro (CLIP-based).
- DA-Ada reaches 66.7% on KITTI to Cityscapes and 67.3% on SIM10k to Cityscapes, the latter above HT's 65.5%.
- DA-Ada reaches 48.0% mAP on Pascal VOC to Clipart, above the best listed baseline CMT at 47.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the source-only adapter with a domain-aware adapter (ablation) | DAOD adapter · Cross-Weather | mAP: 50.4 -> 57.1 | Helped |
| Replace the domain-agnostic adapter with a domain-aware adapter (ablation) | DAOD adapter · Cross-Weather | mAP: 53.8 -> 57.1 | Helped |
| Add DIA to the backbone (ablation) | DA-Ada · Cross-Weather | mAP: 52.6 -> 53.8 | Helped |
| Add the DIA domain adversarial loss (ablation) | DA-Ada with DIA · Cross-Weather | mAP: 53.8 -> 54.8 | Helped |
| Add DSA (ablation) | DA-Ada with DIA and domain adversarial loss · Cross-Weather | mAP: 54.8 -> 56.2 | Helped |
| Add the DIA–DSA decoupling loss (ablation) | DA-Ada with DIA and DSA · Cross-Weather | mAP: 56.2 -> 57.1 | Helped |
| Insert DA-Ada in shallower blocks and increase insertion from one block to four (ablation) | DA-Ada · Cross-Weather | mAP: 55.1 -> 56.9 -> 57.7 -> 58.5 | Helped |
| Set bottleneck dimensions to 32/128/256/512 rather than 16/64/128/256 (ablation) | DA-Ada · Cross-Weather | mAP: 55.9 -> 57.1 | Helped |
| Increase bottleneck dimensions to 48/192/384/768 (ablation) | DA-Ada · Cross-Weather | mAP: 57.1 -> 56.8 | Hurt |
| Use the full input dimensions for the bottleneck (ablation) | DA-Ada · Cross-Weather | mAP: 57.1 -> 56.6 | Hurt |
| Add DITA (ablation) | DA-Ada · Cross-Weather | mAP: 57.1 -> 57.6 | Helped |
| Add the DITA adversarial loss (ablation) | DA-Ada with DITA · Cross-Weather | mAP: 57.6 -> 57.9 | Helped |
Surprising findings
- Domain-agnostic adapter gave similar source-domain performance to the oracle but a 3.7% drop on the target domain, supporting the claim that it is biased toward the source.
- Feeding the raw block output F_i(h) to the DSA gave only limited gains, while feeding the discarded residual h - h^I gave a clear improvement.
Limitations
- Designed only for single labelled source and unlabelled target settings; cannot handle multiple source domains or no target domain.
- Failure cases include missing cars with headlights on in fog and occluded objects, since the source data is mostly sunny.
Experimental setup
Datasets: Cityscapes Foggy Cityscapes KITTI SIM10k Pascal VOC Clipart
Benchmarks: Cityscapes to Foggy Cityscapes KITTI to Cityscapes SIM10k to Cityscapes Pascal VOC to Clipart
Metrics: mAP@0.5
Baselines: DA-Faster SIGMA++ CIGAR CSDA HT D2-UDA AT NSA-UDA DA-Pro CoOp
Builds on: RegionCLIP DA-Pro Faster R-CNN CLIP
Closest papers by meaning
- SEEN-DA: SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection (CVPR 2025 · similarity 0.93)
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (AAAI 2025 · similarity 0.92)
- Differential Alignment for Domain Adaptive Object Detection (AAAI 2025 · similarity 0.90)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.90)
- Non-exemplar Domain Incremental Object Detection via Learning Domain Bias (AAAI 2024 · similarity 0.89)
- DSD-DA: Distillation-based Source Debiasing for Domain Adaptive Object Detection (ICML 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.