Research map / Label-efficient and robust learning / Domain adaptation and generalization
Cross-domain Open-world Discovery
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 CROW, a cluster-then-match prototype method on frozen-then-partially-finetuned CLIP ViT-L features for cross-domain open-world discovery, evaluated across 75 UniDA-derived settings.
Problem
Open-world discovery under domain shift: assign target samples to seen classes and discover novel classes when labeled source and unlabeled target differ in feature distribution.
Approach
Cluster target features from a foundation model first, then robustly match clusters to seen classes (allowing many-to-many matches via column-wise softmax and threshold), treating unmatched clusters as novel; then fine-tune with source cross-entropy plus a target entropy-regularization term.
Key results
- CROW averages 71.1% H-score across 12 seen/unseen splits on four datasets, versus 65.8% for the strongest baseline GLC.
- On the 50% split, CROW reaches 69.6 H-score on OfficeHome, 69.2 on VisDA, and 59.0 on DomainNet, outperforming all baselines.
- Removing the entropy regularization term drops seen-class bias: w/o L_reg gives 77.2 seen / 56.9 unseen versus CROW's 77.0 / 62.8 on VisDA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CROW's cluster-then-match method instead of GLC | Cross-domain open-world discovery · Office, OfficeHome, VisDA, and DomainNet (50% seen/unseen split) | H-score: 74.6 -> 84.9; 68.2 -> 69.6; 65.2 -> 69.2; 56.1 -> 59.0 | Helped |
| Fine-tune both the feature extractor and classifier rather than omit fine-tuning (ablation) | CROW · VisDA (6/6) | H-score: 66.9 -> 69.2 | Helped |
| Fine-tune both the feature extractor and classifier rather than only W (ablation) | CROW · VisDA (6/6) | H-score: 67.8 -> 69.2 | Helped |
| Add supervised loss L_s (ablation) | CROW · VisDA (6/6) | seen accuracy: 65.6 -> 77.0 | Helped |
| Add entropy regularization L_reg (ablation) | CROW · VisDA (6/6) | unseen accuracy: 56.9 -> 62.8 | Helped |
| Use CLIP ViT-L instead of CLIP ViT-B (ablation) | CROW · VisDA (6/6) | H-score: 65.8 -> 69.2 | Helped |
| Scale up DINO_v2 from ViT-B to ViT-G (ablation) | CROW · VisDA (6/6) | H-score: 63.5 -> 68.2 | Helped |
| Scale up SWAG from ViT-B to ViT-H (ablation) | CROW · VisDA (6/6) | H-score: 66.7 -> 70.3 | Helped |
| Exclude unseen-class examples from ImageNet pretraining (ablation) | ResNet-50 · Office | unseen accuracy: 56.4 -> 55.8 | No clear effect |
| Use CROW instead of K-means clustering alone | CLIP features · Office, OfficeHome, and VisDA | H-score: 77.2 -> 84.9; 65.9 -> 69.9; 62.4 -> 69.2 | Helped |
| Use the estimated rather than known number of target classes (ablation) | CROW · Office, OfficeHome, VisDA, and DomainNet (50% split) | H-score: 84.9 -> 81.5; 69.6 -> 67.2; 69.2 -> 68.0; 59.0 -> 57.8 | Hurt |
| Vary λ across the reported values (ablation) | CROW · VisDA (6/6) | H-score: 69.3; 69.0; 68.8; 69.4; 69.2; 69.2 | No clear effect |
Surprising findings
- Baselines were given a test-set-tuned threshold, which the authors note is unrealistic but used to push baseline performance.
- CLIP zero-shot with only the 31 ground-truth class names matches CROW (85.7 vs 84.9 H-score on Office), suggesting the evaluation favors exact vocabulary knowledge.
Limitations
- Performance depends on the clustering quality of the foundation model's feature space; on DomainNet Quickdraw, CROW reaches only ~20-24% accuracy.
- Requires the number of target classes |C_t| as a prior; estimated |C_t| lowers H-score by 1-3 points.
- Baseline thresholds were tuned on the test set, making the comparison optimistic for baselines.
Experimental setup
Datasets: Office OfficeHome VisDA DomainNet
Benchmarks: Office OfficeHome VisDA DomainNet
Metrics: H-score seen accuracy unseen accuracy
Baselines: SIMPLE GCD ORCA DCC DANCE OVANet UniOT NCDDA SAN GLC K-means CLIP zero-shot
Builds on: DCC universal domain adaptation GCD ORCA SimCLR
Closest papers by meaning
- Learning State-Based Node Representations from a Class Hierarchy for Fine-Grained Open-Set Detection (ICML 2025 · similarity 0.84)
- Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.83)
- Discriminative Sample-Guided and Parameter-Efficient Feature Space Adaptation for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.83)
- Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval (NeurIPS 2024 · similarity 0.83)
- Breaking Class Barriers: Efficient Dataset Distillation via Inter-Class Feature Compensator (ICLR 2025 · similarity 0.83)
- CIFAR-10-Warehouse: Broad and More Realistic Testbeds in Model Generalization Analysis (ICLR 2024 · similarity 0.83)
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.