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

Dynamic Target Distribution Estimation for Source-Free Open-Set Domain Adaptation

AAAI 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Feature alignment for unsupervised adaptation · 6 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 DTDE, a source-free open-set domain adaptation method that builds known prototypes via self-adaptive sampling and unknown prototypes via self-splitting k-means clustering with exclusion, plus class-aware KNN alignment.

Problem

Source-free open-set domain adaptation (SF-OSDA): adapting a source-pretrained closed-set model to an unlabeled target domain containing unknown classes, without access to source data, where unknown semantics are unavailable.

Approach

Estimate the target clustering distribution via a greedy self-splitting mechanism (JS divergence-based criterion over k-means clusters), then exclude clusters resembling known-class prototypes to obtain unknown prototypes; combine with self-adaptive high-confidence prototypes for known classes and pseudo-label self-training with class-aware KNN alignment.

Key results

• On VisDA under OSDA, DTDE reaches 80.4% HOS, 7.6 points above the best competitor (GLC/LEAD, around 72.8%).
• DTDE achieves the highest UNK accuracy on all three OSDA benchmarks, e.g. 84.9% UNK on VisDA and 79.8% on Office-Home.
• In the UniDA setting, DTDE attains 79.4% HOS on VisDA, the best among the compared methods.

Measured effects reported in the paper

Surprising findings

• Source-free DTDE matches or surpasses several traditional source-accessible OSDA methods, which the authors frame as a notable result.

Limitations

• Relies on a fixed uncertainty threshold (gamma=0.55) carried over from prior work for known/unknown separation.
• Hyper-parameter alpha must be tuned per dataset (0.3 or 1.5).

Experimental setup

Datasets: Office-31 Office-Home VisDA

Benchmarks: Office-31 Office-Home VisDA

Metrics: HOS UNK accuracy OS*

Baselines: STA OSLPP ROS ANNA SHOT-O GLC LEAD UAN CMU DCC OVANet GATE

Builds on: SHOT GLC ANNA KNN alignment (Qu et al. 2023)

Closest papers by meaning

• Source-Free Domain Adaptation via Distribution Estimation (CVPR 2022 · similarity 0.91)
• Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.91)
• Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.90)
• Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.89)
• Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment (CVPR 2025 · similarity 0.89)
• Open-Set Heterogeneous Domain Adaptation: Theoretical Analysis and Algorithm (AAAI 2025 · similarity 0.88)