Semantic Bridging Domains: Pseudo-Source as Test-Time Connector

Xizhong Yang; Huiming Wang; Ning Xu; Mofei Song

cs.CLMar 4, 2026

Semantic Bridging Domains: Pseudo-Source as Test-Time Connector

Authors: Xizhong Yang, Huiming Wang, Ning Xu, Mofei Song

Organizations: Southeast University · Kuaishou Technology

Abstract

Distribution shifts between training and testing data are a critical bottleneck limiting the practical utility of models, especially in real-world test-time scenarios. To adapt models when the source domain is unknown and the target domain is unlabeled, previous works constructed pseudo-source domains via data generation and translation, then aligned the target domain with them. However, significant discrepancies exist between the pseudo-source and the original source domain, leading to potential divergence when correcting the target directly. From this perspective, we propose a Stepwise Semantic Alignment (SSA) method, viewing the pseudo-source as a semantic bridge connecting the source and target, rather than a direct substitute for the source. Specifically, we leverage easily accessible universal semantics to rectify the semantic features of the pseudo-source, and then align the target domain using the corrected pseudo-source semantics. Additionally, we introduce a Hierarchical Feature Aggregation (HFA) module and a Confidence-Aware Complementary Learning (CACL) strategy to enhance the semantic quality of the SSA process in the absence of source and ground truth of target domains. We evaluated our approach on tasks like semantic segmentation and image classification, achieving a 5.2% performance boost on GTA2Cityscapes over the state-of-the-art.

Figures & tables

Figure 1: Framework of Stepwise Semantic Alignment (SSA). Source Model * and Pre-Trained Model * are their respective feature extractors. TD-S and PRE-S refer to the S emantic features derived from the T arget- D omain (idealized semantics) and the PRE -Trained model, respectively. PS-S and RT-S correspond to the S emantic features extracted from $D_{ps}$ and $D_{rt}$ . See Appendix §A for pseudo-code.

GTA5 $\rightarrow$ Cityscapes (CS)
Method	SF	road	side.	build.	wall	fence	pole	light	sign	vege.	terr.	sky	person	rider	car	truck	bus	train	motor.	bike	mIoU
TransDA-B	✗	94.7	64.2	89.2	48.1	45.8	50.1	60.2	40.8	90.4	50.2	93.7	76.7	47.6	92.5	56.8	60.1	47.6	49.6	55.4	63.9
DAFormer	✗	95.7	70.2	89.4	53.5	48.1	49.6	55.8	59.4	89.9	47.9	92.5	72.2	44.7	92.3	74.5	78.2	65.1	55.9	61.8	68.3
HRDA	✗	96.4	74.4	91.0	61.6	51.5	57.1	63.9	69.3	91.3	48.4	94.2	79.0	52.9	93.9	84.1	85.7	75.9	63.9	67.5	73.8
IDM	✗	97.2	77.1	89.8	51.7	51.7	54.5	59.7	64.7	89.2	45.3	90.5	74.2	46.6	92.3	76.9	59.6	81.2	57.3	62.4	69.5
DAFormer	✓	87.7	33.4	83.9	28.1	27.5	35.9	42.9	28.7	82.4	28.6	83.1	65.0	37.0	85.8	53.9	46.3	31.8	23.6	36.8	49.6

Table 1: Semantic segmentation main results of SSA on different tasks. SF denotes the Source Free method, which does not use source domain data during the alignment process. GTA5, Cityscapes, and ACDC share the same 19 categories, while SYNTHIA has only 16 of them, and its mIoU results are scaled accordingly.

Office-31 (ResNet-50)
Method	SF	A $\to$ D	A $\to$ W	D $\to$ A	D $\to$ W	W $\to$ A	W $\to$ D	Avg
SHOT	✓	93.7	91.1	74.2	98.2	74.6		88.6
DIFO	✓	97.2	95.5	83.0	97.2	83.2	98.8	92.5
ProDe	✓	96.8	96.4	83.1	97.0	82.5	99.8	92.6
SSA	✓	97.0	96.9	82.6	98.2	82.1	99.8	92.8
Office-Home (ResNet-50)

Table 2: Single-Label image classification performance of SSA on different tasks. SF denotes the Source Free method.

Method	HFA	CACL	SSA	mIoU
Baseline				44.5
(a)	✓			57.5
(b)		✓		50.2
(c)			✓	53.4
(d)	✓	✓		65.6
(e)	✓		✓	66.7

Table 3: Ablation Studies on GTA5

\to

Cityscapes.

Panel of Figure 3 — Figure 3: SSA t-SNE visualization on VisDA-C.

Method	Office-Home	VisDA-C	Avg.
TENT	67.27	75.43	71.35
SHOT	67.89	76.69	72.29
SSA	76.62	82.58	79.60

Table 4: Online scenarios results of SSA in on Image Classification.

Task	Method	Ar	Cl	Pr	Rw	Avg.
MTDA	CoNMix	75.6	81.4	71.4	73.4	75.4
	ProDe	83.3	89.2	80.9	81.2	83.6
	SSA	84.1	89.4	81.6	81.6	84.2
MSDA	SHOT-Ens	82.9	82.8	59.3	72.2	74.3
	DECISION	83.6	84.4	59.4	74.5	75.5
	ProDe	91.1	92.5	73.4	83.0	85.0

Table 5: Multi-source/target results of SSA on Office-Home.

Model	$\mathcal{D}_{s}$	$\mathcal{D}_{t}$				Avg.
Model	$\mathcal{D}_{s}$	Ar	Cl	Pr	Rw	Avg.
Pre-Trained	Ar	90.32	33.17	63.62	71.72	64.71
	Cl	44.26	87.39	59.04	54.52	61.30
	Pr	44.68	32.46	91.28	71.38	59.95
	Rw	54.28	35.60	71.57	91.05	63.13
Source	Ar	97.50	44.80	65.88	72.11	70.07

Table 6: Source and pre-trained models results on Office-Home.

Method	COCO2014	COCO2017	VOC2007	VOC2012	mAcc
CLIP	47.53	47.32	75.91	74.25	61.25
TPT	48.52	48.51	75.54	73.92	61.62
ML-TTA	51.58	51.39	78.62	76.63	64.56
SSA	53.27	52.96	79.37	77.29	65.72

Table 7: Multi-label classification of SSA.

Figure 4: Scaling effect of SSA’s gains in different tasks.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Sensitivity analysis for the thresholds.

Method	$\bar{\tau_{\text{par}}}$	Office-31 (0.1335)	Office-Home (0.2491)	VisDA-C (0.1677)	DomainNet (0.3478)
Conf-Aware	0.92	92.8	-	-	-
	0.87	-	85.0	-	-
	0.79	-	-	91.3	-
	0.44	-	-	-	83.7
Fixed	0.10	92.5	84.7	90.9	83.4
	0.20	92.5	84.6	91.2	83.5

Appendix

Table 8: Impact of Different

\tau_{\text{par}}

Parameters on Single-label Classification

$\tau_{\text{lg}}/\ \tau_{\text{lw}}$	GTA5 $\rightarrow$ Cityscapes			VisDA-C
$\tau_{\text{lg}}/\ \tau_{\text{lw}}$	1024×1024	512×512	256×256	256×256	128×128
0.1	68.8	68.9	68.1	91.1	91.3
0.2	68.9	68.8	68.1	91.1	91.4
0.3	69.1	68.9	68.2	91.2	90.9
0.4	69.0	68.7	68.5	91.1	91.0
0.5	69.2	69.1	68.7	91.3	91.1

Appendix

Table 9: Sensitivity analysis of

\tau_{\text{lg}}

with varying

\tau_{\text{lw}}

on different datasets.

Epoch	Baseline	Strategy 1	Strategy 2	Strategy 3
1	88.74	88.90	89.00	89.20
2	89.73	89.91	90.00	90.10
3	90.06	90.15	90.20	90.30
4	90.20	90.31	90.35	90.50
5	90.34	90.47	90.50	90.70
6	90.51	90.63	90.65	90.90

Appendix

Table 10: Performance comparison of mitigation strategies for

\tau_{\text{neg}}

across epochs.

Method	GPU Memory (GB)	Time-Item (s/item)	Time-Train (GPU·h)	mIoU (%)	Backbone	Model Size (MB)
SSA	22.4	2.40	26.7	64.1	Segformer-B5	313.1
SSA w/o CACL	22.4	2.40	26.4	63.4	Segformer-B5	313.1
SSA w/o HFA	20.2	1.52	18.8	61.3	Segformer-B5	313.1
SSA w/o SDA	19.9	2.40	14.3	62.7	Segformer-B5	313.1
SSA-B4	20.4	2.12	22.1	63.5	Segformer-B4	234.4
SSA-B3	17.0	1.84	17.4	62.3	Segformer-B3	170.3

Appendix

Table 11: Performance and Computational Cost Analysis on Semantic Segmentation.

Method	GPU Memory (GB)	Time-Item (s/item)	Time-Train (GPU·h)	mAcc (%)	Backbone	Model Size (MB)
SSA	10.8	0.65	0.17	85.0	ResNet-50	102.6
SSA w/o CACL	10.8	0.65	0.16	84.9	ResNet-50	102.6
SSA w/o HFA	9.7	0.57	0.15	84.6	ResNet-50	102.6
SSA w/o SDA	8.2	0.65	0.13	84.7	ResNet-50	102.6
SSA-RN101	13.6	0.67	0.18	89.3	ResNet-101	178.9
SSA-VGG16	13.9	0.67	0.18	86.7	VGG16	533.4

Appendix

Table 12: Performance and Computational Cost Analysis on Classification.

Method	GPU Memory (GB)	Time-Train (GPU·h)	mIoU/mAcc (%)	Task	Dataset	Backbone
SSA	22.4	26.5	64.1	Segmentation	SYNTHIA $\to$ Cityscapes (0.5396)	Segformer-B5
SSA+MobileViT	21.2	24.2	63.9		SYNTHIA $\to$ Cityscapes (0.5396)
SSA w/o SDA	19.9	14.5	62.7		SYNTHIA $\to$ Cityscapes (0.5396)
SSA	22.4	26.7	69.2		GTA5 $\to$ Cityscapes (0.7239)
SSA+MobileViT	21.2	24.3	68.9		GTA5 $\to$ Cityscapes (0.7239)
SSA w/o SDA	19.9	14.3	65.6		GTA5 $\to$ Cityscapes (0.7239)

Appendix

Table 13: Performance Optimization Results Across Different Tasks.

Method	GPU Memory (GB)	Time-Item (s/item)	Time-Train (GPU $\cdot$ h)	mIoU (%)	Backbone
SSA-B0	11.7 (+0.0)	1.47 (+0.00)	11.2 (+0.0)	48.9 (+0.0)	Segformer-B0
SSA-B1	12.6 (+0.9)	1.50 (+0.03)	11.7 (+0.5)	54.3 (+5.4)	Segformer-B1
SSA-RN101	13.2 (+1.5)	1.55 (+0.08)	11.3 (+0.1)	58.2 (+9.3)	ResNet-101
SSA-P2T	12.3 (+0.6)	1.49 (+0.02)	11.5 (+0.3)	60.2 (+11.3)	P2T-Base
SSA-B2	13.1 (+1.4)	1.66 (+0.09)	12.4 (+1.2)	61.2 (+12.3)	Segformer-B2
SSA-B3	17.0 (+5.3)	1.84 (+0.37)	17.4 (+6.2)	62.3 (+13.4)	Segformer-B3

Appendix

Table 14: Trade-off Analysis Between Computational Cost and Segmentation Performance

Panel of Figure 6 — Figure 6: Confusion matrices for different domain shifts in the Office-Home dataset.

Figure 7: Confusion matrices for different domain shifts in the VisDA-C dataset.

Panel of Figure 8 — Figure 8: Confusion matrices for different domain shifts in the Office-31 dataset.

Panel of Figure 9 — Figure 9: t-SNE visualization for image classification on VisDA-C dataset.

Domain Adaptation

Domain adaptation transfers machine-learning models across changing data distributions, often using unlabeled or limited target-domain examples.

See all 252 papers

LASA: Language-and-Source-Anchored Alignment for Domain Generalized Semantic Segmentation
Aug 9, 2026Jinhong Zhu, Weiqi Yan, Shengchuan Zhang +1Domain Generalization Domain Generalization for Semantic Segmentation
Xiamen University · Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China · Xiamen, Fujian, China
Adaptive Dual-Teacher Distillation with Subnetwork Rectification for Bridging Semantic Gaps in Black-Box Domain Adaptation
Mar 24, 2026Zhe Zhang, Jing Li, Wanli Xue +4Source-Free Domain Adaptation Knowledge Distillation
School of Computer Science and Engineering, Tianjin University of Technology, Tianjin 300384, China. · School of Artificial Intelligence, Tianjin University, Tianjin 300350, China
Vision-Language-Guided Pseudo-Labels for Unsupervised Domain Adaptation in Semantic Segmentation for Waste Sorting
Sep 1, 2026Udo Schlegel, Shubhangi, Gabriel Dax +3Unsupervised Domain Adaptation Text-Guided Image Segmentation
LMU Munich, Munich, Germany · Munich Center for Machine Learning (MCML), Munich, Germany · Fraunhofer Institute for Integrated Circuits IIS, Nuremberg, Germany

Aug 9, 2026J/K move · Enter open · S save

LASA: Language-and-Source-Anchored Alignment for Domain Generalized Semantic Segmentation

Jinhong Zhu, Weiqi Yan, Shengchuan Zhang, Liujuan Cao

Xiamen University · Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China · Xiamen, Fujian, China

Domain Generalization Domain Generalization for Semantic Segmentation Semantic Segmentation

Open paper AskPDF arXiv

Domain Generalization Semantic Segmentation (DGSS) focuses on generalizing knowledge from labeled source domains to unseen target domains where data is unavailable during the training phase. While conventional methods utilize style randomization or feature normalization to mitigate domain shifts, they often impair feature integrity. Specifically, style randomization distorts the underlying feature manifold due to its coarse-grained nature, while feature normalization suppresses discriminative, domain-sensitive semantic details owing to its rigid design. To address these limitations, we propose the Language-and-Source-Anchored Alignment (LASA) framework, which comprises three synergistic components: Text-and-Source-Guided Style Transfer (TSGST), Domain-Aware Query Adapter (DAQA), and Domain-Aware Decoder Optimizer (DADO). Concretely, the TSGST module addresses manifold distortion by utilizing source features as structural anchors and vision-language model (VLM) priors as fine-grained guidance. To restore suppressed discriminative and domain-sensitive details, the DAQA module recalibrates object queries via categorical guidance and domain-aware signatures, while the DADO module aligns the resulting query distributions with a shared classifier to ensure consistent categorical responses across domains. Extensive experiments on challenging benchmarks demonstrate that our method significantly outperforms state-of-the-art approaches.

Abstract

Domain Adaptation

Explore similar work

LASA: Language-and-Source-Anchored Alignment for Domain Generalized Semantic Segmentation

Adaptive Dual-Teacher Distillation with Subnetwork Rectification for Bridging Semantic Gaps in Black-Box Domain Adaptation

Vision-Language-Guided Pseudo-Labels for Unsupervised Domain Adaptation in Semantic Segmentation for Waste Sorting