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

Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment

CVPR 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Conditional distribution alignment · 4 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 DTE, a source-free open set domain adaptation method that first identifies unknown target samples via partial unbalanced optimal transport with weight barcode quantization, then assigns pseudo labels to known samples via sparse sample-label matching with a proximal term.

Problem

Source-free open set domain adaptation, where target data contains unknown classes absent from the source and source data is inaccessible, so known/unknown separation and reliable pseudo-labeling for known samples are hard without prior knowledge of openness.

Approach

Estimates per-sample marginal probabilities via partial unbalanced optimal transport over a range of KL-regularization strengths, binarizes them into a weight barcode and clusters it to flag unknown samples without a threshold; then assigns sparse pseudo labels to known samples using optimal transport with an l2 proximal term, instead of entropic Sinkhorn which yields dense plans.

Key results

• On Office-31 DTE reaches 96.4% average OS and 97.2% average OS*, the best among compared methods (SCDA 95.8/96.8, UOL 95.5/96.2).
• On Office-Home DTE reports 76.5% average OS*, a 2.27-point gain over the runner-up.
• On VisDA-2017 DTE achieves 92.9 OS and 92.7 OS* averages on Digits-to-VisDA-style table entries, and 3.15-point OS* improvement over the runner-up on VisDA-2017 is claimed.

Measured effects reported in the paper

Surprising findings

• DTE-W (replacing weight barcode with JPOT-style mean transport cost) already outperforms most prior methods, and dense Sinkhorn-based self-labelling (DTE-S) is worse than sparse matching.

Limitations

• Relies on hyperparameters (epsilon, tau range) chosen empirically; sensitivity shown only on a grid.
• Evaluation limited to standard image benchmarks with a fixed known/unknown class split.

Experimental setup

Datasets: Office-31 Office-Home VisDA-2017 MNIST USPS SVHN

Benchmarks: Office-31 Office-Home VisDA-2017 Digits

Metrics: OS OS* UNK HOS

Baselines: ResNet DANN DeepJDOT OpenMax OSBP STA TIM JPOT SHOT PGL SFPGL cUADAL

Builds on: SHOT JPOT STA Partial optimal transport

Closest papers by meaning

• Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.91)
• Dynamic Target Distribution Estimation for Source-Free Open-Set Domain Adaptation (AAAI 2025 · similarity 0.89)
• Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.89)
• Unveiling the Unknown: Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation (CVPR 2024 · similarity 0.89)
• COSDA: Counterfactual-based Susceptibility Risk Framework for Open-Set Domain Adaptation (ICML 2025 · similarity 0.88)
• LEAD: Learning Decomposition for Source-free Universal Domain Adaptation (CVPR 2024 · similarity 0.88)