Source-linked AI summary
Source-Free Domain Adaptation via Distribution Estimation
Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, Dacheng Tao
TL;DR
Domain adaptation must transfer knowledge across shifted domains even when privacy constraints prevent access to source data. SFDA-DE estimates class-conditioned source distributions from target data and pretrained classifier anchors, then aligns sampled surrogate features with target features. It reports state-of-the-art SFDA performance on Office-Home and VisDA-2017, including 72.9% and 86.5% average accuracy.
Problem
Domain adaptation requires transferring knowledge from labeled source data to unlabeled target data, but source datasets are often unavailable because of privacy constraints.
Method
SFDA-DE uses anchor-initialized spherical k-means for pseudo-labeling, estimates class-conditioned source distributions from target data, and aligns sampled surrogate features contrastively.
Results
SFDA-DE achieves state-of-the-art performance among SFDA methods on Office-Home at 72.9% and VisDA-2017 at 86.5%.
Takeaways & Limitations
The method can perform source-free adaptation and surpass some traditional domain adaptation methods that require source data.
Takeaways & Limitations
The setting assumes only unlabeled target data are available during adaptation and models source features with class-conditioned multivariate Gaussian distributions.
Abstract
from arXiv · showhide
Domain Adaptation aims to transfer the knowledge learned from a labeled source domain to an unlabeled target domain whose data distributions are different. However, the training data in source domain required by most of the existing methods is usually unavailable in real-world applications due to privacy preserving policies. Recently, Source-Free Domain Adaptation (SFDA) has drawn much attention, which tries to tackle domain adaptation problem without using source data. In this work, we propose a novel framework called SFDA-DE to address SFDA task via source Distribution Estimation. Firstly, we produce robust pseudo-labels for target data with spherical k-means clustering, whose initial class centers are the weight vectors (anchors) learned by the classifier of pretrained model. Furthermore, we propose to estimate the class-conditioned feature distribution of source domain by exploiting target data and corresponding anchors. Finally, we sample surrogate features from the estimated distribution, which are then utilized to align two domains by minimizing a contrastive adaptation loss function. Extensive experiments show that the proposed method achieves state-of-the-art performance on multiple DA benchmarks, and even outperforms traditional DA methods which require plenty of source data.
1. Introduction
Domain adaptation addresses distribution shifts between labeled source and unlabeled target domains, but conventional methods require source data that privacy policies often make unavailable. SFDA-DE addresses this setting by estimating source distributions from target data and pretrained-model knowledge, achieving state-of-the-art results on multiple benchmarks.
- Motivation: CNNs trained in one scenario can suffer severe performance drops when target-domain distributions differ because training and test distributions are assumed identical.Real-world changes such as weather and illumination create these domain shifts.
- Domain adaptation: Traditional domain adaptation transfers knowledge from a labeled source domain to an unlabeled target domain by aligning their feature distributions.Existing methods generally map both domains into a shared feature space.
- Source-free setting: Privacy policies and unreleased training datasets make most traditional domain adaptation methods infeasible because they require source and target data during adaptation.Companies may release pretrained models while retaining the source datasets.
- Source-free setting: Source-Free Domain Adaptation uses only unlabeled target data and a model pretrained on the source domain to transfer knowledge without accessing source data.Prior approaches include information maximization, entropy minimization, and generative target-data enrichment.
- SFDA-DE: SFDA-DE initializes spherical k-means with classifier weight anchors, estimates class-conditioned source distributions from target data, samples surrogate source features, and aligns them contrastively.The method is designed for image classification under SFDA.
- Results: 72.9% on Office-Home and 86.5% on VisDA-2017 are reported as state-of-the-art results among SFDA methods, with superiority to some source-data-dependent methods.The method was evaluated on Office-31, Office-Home, and VisDA-2017.
2. Related Work
Traditional domain adaptation aligns source and target feature distributions, whereas SFDA-DE estimates a surrogate source distribution when source data is unavailable and aligns it with target features.
- Traditional domain adaptation: Traditional domain adaptation methods align source and target feature distributions using techniques such as MMD and contrastive learning.MMD measures differences between probability distributions in an RKHS.
- Source-free domain adaptation: SFDA-DE replaces unavailable source data with an estimated source distribution and uses spherical k-means for more robust pseudo-labeling than classifier-output assignment.The classifier is frozen so its source anchors remain fixed during adaptation.
3. Method
SFDA-DE estimates source distributions without source data through anchor-based pseudo-labeling, class-conditioned distribution estimation, surrogate-feature sampling, and contrastive alignment.
- Method overview: The method obtains robust target pseudo-labels, estimates class-conditioned source distributions, and aligns sampled surrogate source features with target features using contrastive adaptation.These are the method’s three main processing steps.
- Problem setting: SFDA assumes access to a source-pretrained model and unlabeled target data while forbidding access to the source dataset during adaptation.The feature extractor produces m-dimensional representations and the classifier supplies learned class weight vectors.
- Pseudo-labeling: Spherical k-means uses classifier weight vectors as initial class centers, alternates assignment and center updates, and filters ambiguous samples with confidence threshold τ.The strategy addresses classifier degradation caused by target-source distribution shift.
- Source Distribution Estimation: The source feature distribution is approximated with class-conditioned Gaussian distributions whose means are calibrated by anchors and whose covariance statistics are derived from target features.The Gaussian covariance captures within-class feature variation and semantic channel relationships.
- Source Distribution Estimation: Anchor calibration preserves the estimated source mean’s direction while deriving its scale from target features, rather than directly using either anchors or target means.This accounts for empirical norm differences between classifier anchors and feature representations.
- Surrogate source data: The method derives K surrogate source distributions from anchors and target features, then samples surrogate features to simulate unavailable source features.The surrogate distributions are intended to approximate the unknown source distributions.
- Contrastive alignment: Contrastive Domain Discrepancy aligns surrogate source and target distributions by reducing intra-class discrepancy and increasing inter-class discrepancy.Different classes are treated as negative pairs to encourage compactness and separability.
1 Initialize the cluster center with source anchors wG
The training process repeatedly constructs pseudo-labeled target data and surrogate source batches, computes CDD loss, and updates the feature extractor while keeping the classifier fixed.
- Epoch initialization: Each epoch applies spherical k-means to target features and constructs a confident pseudo-labeled set.This precedes surrogate source-distribution estimation.
- Batch construction: SDE derives K surrogate source distributions before target and source mini-batches are sampled for adaptation.The surrogate source batch is sampled from the estimated distributions rather than the unavailable source dataset.
- Optimization: CDD loss is computed from the surrogate source and target batches, then backpropagation updates the feature extractor.The classifier remains frozen during this process.
- Dynamic re-estimation: Pseudo-labels and surrogate distributions are re-estimated every epoch because target features and their statistics change during adaptation.The evolving target features are driven toward the source anchors.
4. Experiments
Experiments across three domain-adaptation benchmarks evaluate SFDA-DE’s performance, hyper-parameter sensitivity, and component effectiveness. Results show strong benchmark performance and support anchor-based pseudo-labeling, dynamic distribution estimation, and surrogate-feature alignment.
- Experimental results: SFDA-DE achieves state-of-the-art performance on Office-Home (72.9%) and VisDA-2017 (86.5%) among SFDA methods.It also outperforms some traditional domain-adaptation methods requiring source data.
- Experimental results: On Office-31, SFDA-DE matches A2Net’s best result (90.1%) and achieves the best average performance and results on 2 of 6 tasks.The dataset’s small class sizes make source-distribution estimation more difficult.
- Experimental results: On VisDA-2017, SFDA-DE exceeds the second-best A2Net by 1.1% and reaches 86.5% average accuracy despite the large synthetic-to-real domain gap.The authors attribute this result to the dataset’s approximately 55K target images for distribution estimation.
- Hyper-parameter sensitivity: For small-scale datasets, increasing τ can lower pseudo-label accuracy while improving performance by including more target data in D′_t.The authors set τ = 0.6 for Office-31 and Office-Home, but τ = 0.078 for VisDA-2017.
- Hyper-parameter sensitivity: The covariance coefficient γ controls surrogate-feature diversity, and γ = 2 improves VisDA-2017 performance by 0.2%.Larger covariance expands the sampling range but an inappropriate value may harm performance.
5. Conclusions
SFDA-DE addresses source-free domain adaptation by estimating source feature distributions without source data and using source anchors to obtain target pseudo-labels.
- SFDA-DE estimates source feature distributions without accessing source data.
- The framework uses domain knowledge preserved by source anchors to obtain high-quality pseudo-labels for target data.
- Experiments validate SFDA-DE’s effectiveness and superiority against other strong SFDA baselines.