Research map / Continual learning / Domain and prompt-based
Non-exemplar Domain Incremental Object Detection via Learning Domain Bias
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 LDB, a non-exemplar domain incremental object detection method that freezes a ViTDet base model after session 1 and trains per-domain MLP/MHA bias tokens, with an NMC-based domain selector at test time.
Problem
Domain incremental object detection (DIOD) requires learning new visual domains sequentially without forgetting old ones, without storing exemplars from previous domains (privacy and storage concerns).
Approach
Freeze the transformer weights after the first domain and learn independent bias-like domain tokens (MLP and MHA biases, modulated by a learned spatial weight) per new domain; at test time a nearest-mean-classifier over ImageNet-pretrained features selects which domain bias to apply, since domain IDs are unknown.
Key results
- On the Pascal VOC series, LDB reaches 68.1%, 64.2%, and 56.8% mAP at sessions 2-4, exceeding the best non-exemplar method CIFRCN by 2.8, 6.5, and 3.3 points.
- On the BDD100K series, LDB reaches 52.3% and 51.1% mAP at sessions 2 and 3, surpassing the exemplar-based S-Prompt by 0.7 and 1.7 points.
- Prompt-based comparison at session 4 on VOC: LDB 56.8% mAP with 0.19% extra parameters vs. L2P 45.5% and S-Prompt 45.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add domain-specific bias and predictors, with NMC domain selection, rather than exemplar-based TP-DIOD-B | ViTDet on Pascal VOC series · Pascal VOC series, session 2 | mAP: 65.8 -> 68.1 | Helped |
| Add domain-specific bias and predictors, with NMC domain selection, rather than exemplar-based TP-DIOD-B | ViTDet on Pascal VOC series · Pascal VOC series, session 4 | mAP: 57.5 -> 56.8 | Hurt |
| Use LDB instead of L2P | ViTDet on Clipart · Clipart, session 4 | AP: 32.5 -> 50.1 | Helped |
| Use LDB instead of L2P | ViTDet on Watercolor · Watercolor, session 4 | AP: 43.3 -> 57.5 | Helped |
| Use LDB instead of S-Prompt | ViTDet on Cityscape · Cityscape, session 3 | AP: 52.0 -> 52.7 | Helped |
| Add MLP bias to the FT-FC baseline (ablation) | ViTDet on Pascal VOC series · Pascal VOC series, session 4 | mAP: 44.2 -> 55.3 | Helped |
| Add MHA bias to the FT-FC baseline (ablation) | ViTDet on Pascal VOC series · Pascal VOC series, session 4 | mAP: 44.2 -> 55.6 | Helped |
| Use NMC instead of K&K with both MLP and MHA biases (ablation) | ViTDet on Pascal VOC series · Pascal VOC series, session 4 | mAP: 55.9 -> 56.8 | Helped |
| Add domain bias to blocks 7–12 instead of blocks 10–12 (ablation) | ViTDet on Pascal VOC series · Pascal VOC series, session 4 | mAP: 51.4 -> 55.2 | Helped |
| Add domain bias to blocks 4–12 instead of blocks 10–12 (ablation) | ViTDet on Pascal VOC series · Pascal VOC series, session 4 | mAP: 51.4 -> 56.6 | Helped |
| Add domain bias to all transformer blocks (ablation) | ViTDet on Pascal VOC series · Pascal VOC series, session 4 | mAP: 51.4 -> 56.8 | Helped |
Surprising findings
- Setting the ViT bias to zero on MNIST gave 42.03% accuracy versus 39.07% with bias, indicating weights alone cluster samples but bias is needed for domain adaptation (motivating observation).
- Standard SFDA methods (MCC, IRG) perform poorly on DIOD even when new-domain labels are provided, with MCC dropping to 23.7% mAP at session 4.
Limitations
- Domain selection relies on NMC over ImageNet-pretrained features, so misrouting of test images to the wrong domain bias is possible and not analyzed in depth.
- Evaluation is limited to two dataset series with a small number of domains (3-4 sessions) and shared object categories.
Experimental setup
Datasets: Pascal VOC 2007 Clipart Watercolor Comic BDD100K Cityscape Rainy Cityscape ImageNet-1K SVHN MNIST
Benchmarks: Pascal VOC series (VOC2007, Clipart, Watercolor, Comic) BDD100K series (BDD100K, Cityscape, Rainy Cityscape)
Metrics: mAP@0.5
Baselines: FT-seq FT-FC MCC IRG LwF PASS L2P S-Prompt CIFRCN ERD TP-DIOD-B
Builds on: ViTDet BitFit AdapterBias S-Prompt TP-DIOD
Closest papers by meaning
- DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.89)
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (AAAI 2025 · similarity 0.89)
- Triple Feature Disentanglement for One-Stage Adaptive Object Detection (AAAI 2024 · similarity 0.88)
- SEEN-DA: SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection (CVPR 2025 · similarity 0.88)
- CAT: Exploiting Inter-Class Dynamics for Domain Adaptive Object Detection (CVPR 2024 · similarity 0.88)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.88)
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.