Research map / Label-efficient and robust learning / Domain adaptation and generalization

Unveiling the Unknown: Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation

CVPR 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Conditional distribution alignment · 8 citations (OpenAlex)

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 an unknown diffuser that infers the target class number via Metropolis-Hastings split/merge over clusters, then uses contrastive constraints and high-confidence pseudo-labels over the explored wider class space for open-set source-free domain adaptation.

Problem

Open-set source-free domain adaptation, where the target domain has extra unknown classes and only a source pre-trained model is available; existing methods group all unknown classes into one, limiting known-class transfer and unknown-class generalization.

Approach

Explore the target class space by dynamically inferring the number of target classes via an MLP-based unknown diffuser that splits and merges clusters with a Metropolis-Hastings acceptance step, then use reliable known-class pseudo-labels and clustered pseudo-labels over the wider space as supervision.

Key results

• On Office-Home, the method reaches 61.6% OS*, 79.7% UNK and 69.2% HOS on average, versus AaD's 59.4%-ish OS* (reported as 3.6% relative gain) and HOS gains of 5.6% relative over AaD.
• On Office-31, the method improves average OS* by 10.4% relative and HOS by 2.3% relative over AaD.
• Its inferred target class numbers and ACC on Office-Home tasks exceed FINCH, DenMune and COMIC on all tasks.

Measured effects reported in the paper

Limitations

• Relies on a manually-set confidence threshold for selecting high-confidence known samples
• Performance on Office-31 is less dominant than on Office-Home; OS* is lower than SHOT on some indicators
• VisDA-C results are only in supplementary material and not in the main text

Experimental setup

Datasets: Office-31 Office-Home

Benchmarks: Office-31 Office-Home

Metrics: OS* UNK HOS ACC inferred class number K

Baselines: Cluster (ResNet-50) SHOT AaD FINCH DenMune COMIC

Builds on: SHOT AaD FINCH Metropolis-Hastings

Closest papers by meaning

• Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment (CVPR 2025 · similarity 0.89)
• Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.88)
• Dynamic Target Distribution Estimation for Source-Free Open-Set Domain Adaptation (AAAI 2025 · similarity 0.88)
• Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.87)
• Adapting to Distribution Shift by Visual Domain Prompt Generation (ICLR 2024 · similarity 0.87)
• Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.86)