ProCal: Probability Calibration for Neighborhood-Guided Source-Free Domain Adaptation

Ying Zheng; Yiyi Zhang; Yi Wang; Lap-Pui Chau

cs.CVMar 19, 2026

ProCal: Probability Calibration for Neighborhood-Guided Source-Free Domain Adaptation

Authors: Ying Zheng, Yiyi Zhang, Yi Wang, Lap-Pui Chau

Organizations: Department of Electrical and Electronic Engineering, The Hong Kong Polytechnic University, Hong Kong, China. · Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong, China.

Abstract

Source-Free Domain Adaptation (SFDA) adapts pre-trained models to unlabeled target domains without requiring access to source data. Although state-of-the-art methods leveraging local neighborhood structures show promise for SFDA, they tend to over-rely on prediction similarity among neighbors. This over-reliance accelerates the forgetting of source knowledge and increases susceptibility to local noise overfitting. To address these issues, we introduce ProCal, a probability calibration method that dynamically calibrates neighborhood-based predictions through a dual-model collaborative prediction mechanism. ProCal integrates the source model's initial predictions with the current model's online outputs to effectively calibrate neighbor probabilities. This strategy not only mitigates the interference of local noise but also preserves the discriminative information from the source model, thereby achieving a balance between knowledge retention and domain adaptation. Furthermore, we design a joint optimization objective that combines a soft supervision loss with a diversity loss to guide the target model. Our theoretical analysis shows that ProCal converges to an equilibrium where source knowledge and target information are effectively fused, reducing both knowledge forgetting and overfitting. We validate the effectiveness of our approach through extensive experiments on 31 cross-domain tasks across four public datasets. Our code is available at: https://github.com/zhengyinghit/ProCal.

Figures & tables

Fig. 1: Illustration of the effect of neighborhood distribution calibration. (a) Direct optimization with neighborhood constraints may lead to source knowledge forgetting and overfitting to noisy local target structures. (b) Neighborhood distribution calibration mitigates these issues and yields better generalization.

Method	Sup.	Nei.	Cal.	Off.-31	Off.-Home	VisDA	DomainNet
SHOT [ 15 ]	hard	✗	✗	88.6	71.8	82.9	68.9
AaD [ 19 ]	soft	✓	✗	89.9	72.7	88.0	70.5
ProCal (ours)	soft	✓	✓	90.7	73.8	88.6	72.8

TABLE I: Feature-accuracy (%) comparison of representative SFDA methods. Sup., Nei., Cal., and Off. denote Supervision, Neighborhood, Calibration, and Office, respectively. Leveraging the proposed probability calibration, we achieve consistent improvements across four public datasets.

Fig. 2: Training dynamics on the Ar

\rightarrow

Cl task of Office-Home. Compared with SHOT and AaD, ProCal consistently exhibits a lower source forgetting rate and a lower incorrect supervision rate throughout training, while achieving the highest target accuracy. This suggests that ProCal better preserves transferable source knowledge and reduces harmful target supervision during adaptation.

Method	Ar $\rightarrow$ Cl	Ar $\rightarrow$ Pr	Ar $\rightarrow$ Rw	Cl $\rightarrow$ Ar	Cl $\rightarrow$ Pr	Cl $\rightarrow$ Rw	Pr $\rightarrow$ Ar	Pr $\rightarrow$ Cl	Pr $\rightarrow$ Rw	Rw $\rightarrow$ Ar	Rw $\rightarrow$ Cl	Rw $\rightarrow$ Pr	Avg.
SHOT [ 15 ]	57.1	78.1	81.5	68.0	78.2	78.1	67.4	54.9	82.2	73.3	58.8	84.3	71.8
G-SFDA [ 31 ]	57.9	78.6	81.0	66.7	77.2	77.2	65.6	56.0	82.2	72.0	57.8	83.4	71.3
NRC [ 18 ]	57.7	80.3	82.0	68.1	79.8	78.6	65.3	56.4	83.0	74.0	58.6	85.6	72.2
SHOT++ [ 22 ]	57.9	79.7	82.5	68.5	79.6	79.3	68.5	57.0	83.0	73.7	60.7	84.9	73.0
CoWA-JMDS [ 45 ]	56.9	78.4	81.0	69.1	80.0	79.9	67.7	57.2	82.4	72.8	60.5	84.5	72.5
AaD [ 19 ]	59.3	79.3	82.1	68.9	79.8	79.5	67.2	57.4	83.1	72.1	58.5	85.4	72.7

TABLE II: Classification accuracies (%) using ResNet-50 on the Office-Home dataset. The boldface denotes top performance, while underlined values indicate runner-up results.

Method	C $\rightarrow$ P	C $\rightarrow$ R	C $\rightarrow$ S	P $\rightarrow$ C	P $\rightarrow$ R	P $\rightarrow$ S	R $\rightarrow$ C	R $\rightarrow$ P	R $\rightarrow$ S	S $\rightarrow$ C	S $\rightarrow$ P	S $\rightarrow$ R	Avg.
SHOT [ 15 ]	63.0	78.1	61.1	68.0	81.2	62.6	69.5	68.9	60.2	70.9	65.5	78.0	68.9
G-SFDA [ 31 ]	64.7	77.6	62.8	70.9	81.5	67.2	71.0	70.6	62.9	73.9	68.1	78.1	70.8
NRC [ 18 ]	43.0	40.5	46.5	48.6	49.2	46.5	49.0	51.3	41.5	55.7	50.1	44.6	47.2
DaC [ 23 ]	65.4	79.9	62.9	64.7	81.7	64.0	66.4	69.8	61.7	68.5	67.1	78.2	69.2
CoWA-JMDS [ 45 ]	66.8	81.6	63.5	70.7	82.2	65.7	71.6	69.9	64.7	74.0	68.7	81.3	71.7
AaD [ 19 ]	64.6	79.0	61.7	69.7	82.2	65.9	71.3	70.7	62.2	71.5	67.8	79.2	70.5

TABLE III: Classification accuracies (%) using ResNet-50 on the DomainNet-126 dataset.

Method	plane	bcycl	bus	car	horse	knife	mcycl	person	plant	skbrd	train	truck	Per-class
SHOT [ 15 ]	94.3	88.5	80.1	57.3	93.1	94.9	80.7	80.3	91.5	89.1	86.3	58.2	82.9
G-SFDA [ 31 ]	96.1	83.3	85.5	74.1	97.1	95.4	89.5	79.4	95.4	92.9	89.1	42.6	85.4
NRC [ 18 ]	96.8	91.3	82.4	62.4	96.2	95.9	86.1	80.6	94.8	94.1	90.4	59.7	85.9
SHOT++ [ 22 ]	97.7	88.4	90.2	86.3	97.9	98.6	92.9	84.1	97.1	92.2	93.6	28.8	87.3
CoWA-JMDS [ 45 ]	96.2	89.7	83.9	73.8	96.4	97.4	89.3	86.8	94.6	92.1	88.7	53.8	86.9
AaD [ 19 ]	97.4	90.5	80.8	76.2	97.3	96.1	89.8	82.9	95.5	93.0	92.0	64.7	88.0

TABLE IV: Classification accuracies (%) using ResNet-101 on the VisDA-C dataset (Synthesis

\rightarrow

Real).

Method	A $\rightarrow$ D	A $\rightarrow$ W	D $\rightarrow$ A	D $\rightarrow$ W	W $\rightarrow$ A	W $\rightarrow$ D	Avg.
SHOT [ 15 ]	94.0	90.1	74.7	98.4	74.3	99.9	88.6
NRC [ 18 ]	96.0	90.8	75.3	99.0	75.0		89.4
SHOT++ [ 22 ]	94.3	90.4	76.2	98.7	75.8	99.9	89.2
CoWA-JMDS [ 45 ]	94.4	95.2	76.2	98.5	77.6	99.8	90.3
AaD [ 19 ]	96.4	92.1	75.0	99.1	76.5		89.9
NRC+ELR [ 29 ]	93.8	93.3	76.2	98.0	76.9		89.6

TABLE V: Classification accuracies (%) on Office-31.

Method	Ar $\rightarrow$ Cl	Ar $\rightarrow$ Pr	Ar $\rightarrow$ Rw	Cl $\rightarrow$ Ar	Cl $\rightarrow$ Pr	Cl $\rightarrow$ Rw	Pr $\rightarrow$ Ar	Pr $\rightarrow$ Cl	Pr $\rightarrow$ Rw	Rw $\rightarrow$ Ar	Rw $\rightarrow$ Cl	Rw $\rightarrow$ Pr	Avg.
Partial-set
Source	45.2	70.4	81.0	56.2	60.8	66.2	60.9	40.1	76.2	70.8	48.5	77.3	62.8
SHOT [ 15 ]	64.8	85.2	92.7	76.3	77.6	88.8	79.7	64.3	89.5	80.6	66.4	85.8	79.3
HCL [ 58 ]	65.6	85.2	92.7	77.3	76.2	87.2	78.2	66.0	89.1	81.5	68.4	87.3	79.6
AaD [ 19 ]	67.0	83.5	93.1	80.5	76.0	87.6	78.1	65.6	90.2	83.5	64.3	87.3	79.7
CRS [ 59 ]	68.6	85.1	90.9	80.1	79.4	86.3	79.2	66.1	90.5	82.2	69.5	89.3	80.6

TABLE VI: Classification accuracies (%) on the Office-Home datasets under partial-set and open-set settings (ResNet-50 backbone).

#	$\mathcal{L}_{\text{soft}}$	$\mathcal{L}_{\text{div}}$	Office-31	Office-Home	Avg.
1	Source model only		79.2	60.2	69.7
2	✓	✗	85.4	66.9	76.2
3	✗	✓	76.0	39.8	57.9
4	✓	✓	90.7	73.8	82.3
5	$\mathcal{L}_{\text{soft}}$ w/o target ( $\bm{p}_{i}^{t}$ )		90.2	72.9	81.6
6	$\mathcal{L}_{\text{soft}}$ w/o source ( $\bm{p}_{i}^{s}$ )		90.3	73.3	81.8

TABLE VII: Component-wise ablation study (%).

Backbone	Source	SHOT [ 15 ]	AaD [ 19 ]	ProCal	ProCal++
ResNet-50 [ 57 ]	60.2	71.8	72.7	73.8	74.1
ViT-B/16 [ 60 ]	68.9	75.5	76.4	77.4	78.7
ConvNeXt-T [ 61 ]	69.6	77.9	78.5	79.0	80.8

TABLE VIII: Average accuracy (%) of SFDA methods on Office-Home with different backbone architectures.

Method	Office-31	Office-Home	VisDA-C	DomainNet-126
ProCal	90.7 $\pm$ 0.12	73.8 $\pm$ 0.06	88.6 $\pm$ 0.18	72.8 $\pm$ 0.08
ProCal++	90.8 $\pm$ 0.07	74.1 $\pm$ 0.05	88.9 $\pm$ 0.05	73.4 $\pm$ 0.11

TABLE IX: Mean accuracy (%) and standard deviation over repeated runs on four SFDA benchmarks. The consistently small variances indicate that the performance gains of ProCal and ProCal++ are stable across different runs.

Panel of Figure 12 — Fig. 3: Feature space visualizations of the representations learned by (a) the source model, (b) AaD [ 19 ] , and (c) our ProCal on the Ar $\rightarrow$ Cl task of Office-Home. ProCal produces more compact and better-separated class clusters.

Task group	#Pairs	#Top-2 results of ProCal
SOTA–Oracle gap $>5\%$	20	18
SOTA–Oracle gap $\leq 5\%$	11	3

TABLE X: Task-level analysis of gain robustness across 31 domain pairs, grouped by the gap between previous SOTA and Oracle performance.

Method	$\eta=0.2$	$\eta=0.4$	$\eta=0.6$	$\eta=0.8$
Source model	48.5	36.7	24.6	12.7
ProCal	73.4	73.2	72.8	72.5

TABLE XI: Robustness to inaccurate source priors on Office-Home under different noise rates

\eta

injected into source-model predictions. Higher

\eta

indicates more severely corrupted priors.

Fig. 4: Sensitivity analysis of the hyperparameters

\gamma_{1}

\beta_{1}

, and

k

on Office-31, VisDA-C, and Office-Home. Although the performance varies to some extent with different hyperparameter choices, it remains relatively stable in the vicinity of the selected values, suggesting that the proposed method is not overly sensitive to hyperparameter tuning.

$\tau$	Office-31	Office-Home	VisDA-C
1	90.3	73.7	76.2
2	90.6	73.1	78.9
4	90.5	72.8	79.7
6	90.0	72.4	79.8
8	89.8	72.0	88.1
10	N/A	71.7	88.6

TABLE XII: Effect of the memory update frequency

\tau

on the classification accuracy (%) of ProCal on Office-31, Office-Home, and VisDA-C.

Source-Free Domain Adaptation

Source-Free Domain Adaptation transfers models to new data distributions without access to the original training examples.

See all 69 papers

Jul 31, 2026

Domain-Division based Progressive Learning for Source-Free Domain Adaptation

With growing privacy and portability concerns, source-free domain adaptation requires only a source pre-trained model and an unlabeled target domain, allowing for effective adaptation to the target data. Most existing self-training methods focus on selecting and exploiting samples with reliable predictions, often neglecting others. Inspired by the finding that deep models learn clean samples faster than noisy ones, we propose a domain-division based progressive learning method named DPL. Specifically, our approach consists of two alternating stages, each beginning with the division of the target domain into easy-to-adapt and hard-to-adapt subdomains based on adaptation difficulty, followed by neighborhood-based pseudo label assignment. In stage one, we enhance classification accuracy through uncertainty-aware self-training and alignment of corresponding classes between subdomains. Stage two then applies tailored learning strategies to each subdomain, starting with consistency learning on the easy-to-adapt samples and progressing to utilizing local structural information for the more challenging ones, thereby mining the intrinsic properties of the target data. Extensive experiments on several widely used benchmarks validate the effectiveness of our approach, demonstrating superior performance compared to state-of-the-art methods. Our code is available at https://github.com/iamjingli/DPL.

Pan Liu, Jing Li, Meng Zhao +3

School of Computer Science and Engineering, Tianjin University of Technology, Tianjin 300384, China

Source-Free Domain Adaptation Unsupervised Domain Adaptation Self-Training Pseudo-Labeling Curriculum Learning

Ask PDF

Oct 2, 2026

Safeguarding Mutual Correction in Source-Free Domain Adaptation via Cut Statistics

Source-Free Domain Adaptation (SFDA) aims to adapt a source-pretrained model to an unlabeled target domain without access to the original source domain. While early single-model approaches rely on self-refinement, they are inherently susceptible to confirmation bias and struggle to correct their own systematic errors. To overcome this limitation, recent methods introduce Vision-Language (ViL) models as external knowledge sources. However, these approaches operate in a largely unidirectional paradigm, using the ViL model primarily to supervise the source-pretrained model. This overlooks a key structural property: the two models exhibit distinct failure modes -- where one produces an incorrect prediction, the other may produce a correct one, creating a natural opportunity for mutual correction within the target domain. Yet, without ground-truth labels, identifying which model is correct on any given sample is non-trivial, and naively exchanging predictions risks propagating errors across models. To address this challenge, we propose SafeCut, a novel approach that leverages the cut statistic as a label-free measure of prediction reliability to gate cross-model supervision. Our approach dynamically controls both the direction and strength of supervision based on relative reliability, selectively amplifying true corrections while suppressing miscorrections on a per-sample basis. We further provide theoretical justification showing that this reliability-gated mechanism guarantees a net-positive correction signal. Extensive experiments across diverse SFDA benchmarks demonstrate that SafeCut achieves state-of-the-art performance, highlighting the effectiveness of safeguarding mutual correction in SFDA via cut statistics.

Seongjun Lee, Changhee Lee

Department of Artificial Intelligence Korea University Seoul, Republic of Korea

Source-Free Domain Adaptation VLM Adaptation

Ask PDF

May 4, 2026

Rethinking the Need for Source Models: Source-Free Domain Adaptation from Scratch Guided by a Vision-Language Model

Source-Free Domain Adaptation (SFDA) adapts source models to target domains without accessing source data, addressing privacy and transmission issues. However, existing methods still initialize from a source pre-trained model and thus are not truly source-free. Recent works have introduced Vision-Language (ViL) models to guide the adaptation process, in these methods, we observe that for the same target domain, different source models yield minimal variation in final results, indicating the source model itself has limited impact. Motivated by this, we propose ViL-Only Domain Adaptation (VODA) , a stricter setting that eliminates all dependencies on source domain, relying solely on a randomly initialized model, a ViL model, and unlabeled target data. We analyze the adaptation dynamics of VODA and introduce Two-Stage Denoised-Region Distillation (TS-DRD) , a two-stage framework that first warms up the model with ViL guidance, then seek a Denoised-Region inherent in both the ViL and adapting model, yielding cleaner supervision for distillation. Experiments on Office-Home, VisDA, and DomainNet-126 show that under VODA, TS-DRD achieves competitive or superior performance to existing SFDA methods that still use source models, demonstrating its effectiveness and the potential of the VODA setting.

Zhou Bingtao, Xiang Mian, Ning Qian

Sichuan University, No.24 South Section 1, Yihuan Road, Chengdu, 610065, Sichuan, China · Hubei Minzu University, 39 Xueyuan Road, Enshi, 445000, Hubei, China

Source-Free Domain Adaptation VLM Adaptation

Ask PDF

Abstract

Source-Free Domain Adaptation

Explore similar work

Domain-Division based Progressive Learning for Source-Free Domain Adaptation

Safeguarding Mutual Correction in Source-Free Domain Adaptation via Cut Statistics

Rethinking the Need for Source Models: Source-Free Domain Adaptation from Scratch Guided by a Vision-Language Model