Research map / Label-efficient and robust learning / Domain adaptation and generalization
PADCLIP: Pseudo-labeling with Adaptive Debiasing in CLIP for Unsupervised Domain Adaptation
Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, Chen-Nee Chuah
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune CLIP naively on source domain (ablation) | CLIP ResNet-101 · VisDA-2017 target val | accuracy (%): 82.3% -> 55.6% | Hurt |
| add Catastrophic Forgetting Measurement (CFM) to adjust learning rate (ablation) | CLIP ResNet-101 fine-tuning · VisDA-2017 target val | accuracy (%): 55.6% -> 83.9% | Improved |
| add pseudo-labeling on target domain (ablation) | CLIP + CFM · VisDA-2017 target val | accuracy (%): 83.9% -> 86.0% | Improved |
| add adaptive debiasing (CFM and DCM) to pseudo-labeling (ablation) | CLIP + CFM + PL · VisDA-2017 target val | accuracy (%): 86.0% -> 88.5% | Improved |
| add CFM to DebiasPL supervised loss (ablation) | DebiasPL · VisDA-2017 target val | accuracy (%): 64.4% -> 85.9% | Improved |
| add domain name to prompt (ablation) | DebiasPL + CFM · VisDA-2017 target val | accuracy (%): 85.9% -> 86.5% | Improved |
| add Dual Consistency Momentum to pseudo-label debiasing (ablation) | DebiasPL + CFM · VisDA-2017 target val | accuracy (%): 87.8% -> 88.5% | Improved |
| reduce pretraining data from 400M to 3M pairs (ablation) | CLIP · VisDA-2017 target val | accuracy (%): 82.3% -> 59.1% | Hurt |
| lower learning rate 50x or early stop instead of CFM (ablation) | CLIP ResNet-101 · VisDA-2017 target val | accuracy (%): 55.6% -> 75.8% | Improved |
| apply method to ResNet-101 with ImageNet-1K pretraining | PADCLIP · VisDA-2017 | accuracy | No clear effect |
| apply CFM to incremental learning and domain generalization | CLIP fine-tuning · incremental learning, domain generalization | accuracy: +5.3%, +2.4% | Improved |
Closest papers by meaning
- HVCLIP: High-dimensional Vector in CLIP for Unsupervised Domain Adaptation (ECCV 2024 · similarity 0.90)
- Learning to Adapt Frozen CLIP for Few-Shot Test-Time Domain Adaptation (ICLR 2025 · similarity 0.88)
- Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.87)
- Targeted Forgetting of Image Subgroups in CLIP Models (CVPR 2025 · similarity 0.87)
- OSLoPrompt: Bridging Low-Supervision Challenges and Open-Set Domain Generalization in CLIP (CVPR 2025 · similarity 0.86)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 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.
