Research map / Label-efficient and robust learning / Domain adaptation and generalization
Target Semantics Clustering via Text Representations for Robust Universal Domain Adaptation
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.
Searches for target semantic centers as CLIP text embeddings from class names and WordNet nouns via greedy search, then refines CLIP with LoRA using information maximization, plus a UniMS scoring function for unknown detection in universal domain adaptation.
Problem
Universal domain adaptation must identify common-class samples and align them under unknown domain and category shift, but clustering centers in unconstrained continuous image space are domain-biased and have uncontrolled granularity, making the number of target clusters hard to estimate.
Approach
Constrain target semantic centers to a discrete text-embedding space (source class names plus WordNet nouns) searched greedily with frozen CLIP encoders, with a binary gate that learns the number of clusters; then fine-tune CLIP with LoRA using information maximization, and score unknown samples with UniMS weighted by source/target entropy and a GMM threshold.
Key results
- On Office-Home, the method reaches 85.4 average H-score / accuracy across four category shift scenarios, versus 80.9 for the best CLIP-based baseline UniOT*.
- On VisDA it reports 90.4% accuracy, 11.3 points above the best comparison method.
- On DomainNet it averages 80.5/69.7/69.1/69.3/69.6/81.5/73.3 across tasks, exceeding existing methods by 6.0 points on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TASC loss and UniMS to CE training (ablation) | CLIP · Office OPDA | H-score: 82.9 -> 91.3 | Helped |
| add TASC loss to CE training with UniMS (ablation) | CLIP · Office ODA | H-score: 95.2 -> 96.1 | Helped |
| use estimated target-private class proportions for GMM optimization (ablation) | UniMS unknown detection · Office-Home OPDA | H-score: 84.9 -> 89.4 | Helped |
| add target-semantic similarity to the source-semantic score in UniMS (ablation) | unknown detection scoring · Office-Home OPDA | AUROC: 95.1 -> 96.4 | Helped |
| weight source similarity by source entropy (ablation) | MS-s unknown detection score · Office-Home OPDA | AUROC: 95.1 -> 95.9 | Helped |
| use TASC via text representations instead of the strongest reported CLIP baseline | CLIP UniDA · DomainNet OPDA | H-score: 67.3 -> 73.3 | Helped |
| use TASC via text representations instead of the strongest reported CLIP baseline | CLIP UniDA · VisDA SR OPDA | H-score: 79.1 -> 90.4 | Helped |
Surprising findings
- The optimal lambda_div (0.6) is lower than the value used in SHOT-IM, attributed to partially undiscarded private classes and over-clustering.
Limitations
- Requires a predefined upper bound K0 on the number of clusters (100 for Office, 400 for DomainNet).
- Depends on the quality of the WordNet noun vocabulary for representing target-private classes.
- Evaluated only with ViT-B/16 CLIP backbone and LoRA rank 8.
Experimental setup
Datasets: Office Office-Home VisDA DomainNet WordNet nouns
Benchmarks: Office Office-Home DomainNet VisDA
Metrics: H-score classification accuracy AUROC NMI
Baselines: SO DCC DANCE OVANet GATE GLC UniOT
Builds on: CLIP SHOT-IM MLS LoRA UniOT
Closest papers by meaning
- Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.90)
- Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.90)
- Unified Language-driven Zero-shot Domain Adaptation (CVPR 2024 · similarity 0.90)
- Language-Assisted Image Clustering Guided by Discriminative Relational Signals and Adaptive Semantic Centers (NeurIPS 2026 · similarity 0.88)
- Semantic Library Adaptation: LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation (CVPR 2025 · similarity 0.88)
- Tackling Dimensional Collapse toward Comprehensive Universal Domain Adaptation (ICML 2025 · 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.