CGSA: Class-Guided Slot-Aware Adaptation for Source-Free Object Detection

Boyang Dai; Zeng Fan; Zihao Qi; Meng Lou; Yizhou Yu

cs.CVFeb 26, 2026

CGSA: Class-Guided Slot-Aware Adaptation for Source-Free Object Detection

Authors: Boyang Dai, Zeng Fan, Zihao Qi, Meng Lou, Yizhou Yu

Organizations: School of Computing and Data Science, The University of Hong Kong · Department of Electrical & Computer Engineering, National University of Singapore · School of Information and Electronics, Beijing Institute of Technology

Abstract

Source-Free Domain Adaptive Object Detection (SF-DAOD) aims to adapt a detector trained on a labeled source domain to an unlabeled target domain without retaining any source data. Despite recent progress, most popular approaches focus on tuning pseudo-label thresholds or refining the teacher-student framework, while overlooking object-level structural cues within cross-domain data. In this work, we present CGSA, the first framework that brings Object-Centric Learning (OCL) into SF-DAOD by integrating slot-aware adaptation into the DETR-based detector. Specifically, our approach integrates a Hierarchical Slot Awareness (HSA) module into the detector to progressively disentangle images into slot representations that act as visual priors. These slots are then guided toward class semantics via a Class-Guided Slot Contrast (CGSC) module, maintaining semantic consistency and prompting domain-invariant adaptation. Extensive experiments on multiple cross-domain datasets demonstrate that our approach outperforms previous SF-DAOD methods, with theoretical derivations and experimental analysis further demonstrating the effectiveness of the proposed components and the framework, thereby indicating the promise of object-centric design in privacy-sensitive adaptation scenarios. Code is released at https://github.com/Michael-McQueen/CGSA.

Figures & tables

Figure 1: Motivation of the proposed CGSA . (a) Most popular methods focus on filtering pseudo labels. (b) Our CGSA proposes a slot-aware framework to serve as a bridge for aligning common object-level structural features without accessing source data.

Figure 2: Framework of CGSA . It consists of two stages: source-domain pretraining and target-domain adaptation, which work collaboratively to improve SF-DAOD performance.

Figure 3: (a) Pipeline of the proposed HSA module. Through the hierarchical design, the features are first decomposed into coarse-to-fine slots. After projection, these slots are concatenated with the object queries to form slot-aware queries, thereby providing object-level structural priors. (b) Pipeline of the proposed CGSC module. By maintaining the global class prototype and assigning pseudo labels to weighted slots for class attributes, contrastive learning is used to implicitly supervise and guide the slots to focus on domain-invariant yet class-relevant object features.

Method	SF	Base	Person	Rider	Car	Truck	Bus	Mcycle	Bicycle	mAP
Oracle	-	DETR	68.2	53.2	84.0	67.7	67.9	54.1	54.7	64.3
DA-Faster ( Chen et al., 2018 )	✗	FRCNN	28.9	27.4	44.2	19.1	18.0	14.2	22.4	24.9
SFA ( Wang et al., 2021 )	✗	DETR	40.2	27.6	57.5	19.1	23.4	15.4	19.2	28.9
AQT ( Huang et al., 2022 )	✗	DETR	38.2	33.0	58.4	17.3	18.4	16.9	23.5	29.4
MRT ( Zhao et al., 2023 )	✗	DETR	48.4	30.9	63.7	24.7	25.5	20.2	22.6	33.7
DATR ( Chen et al., 2025 )	✗	DETR	58.5	42.8	73.4	26.9	39.9	24.2	37.3	43.3

Table 1: Adaptation from Small-Scale to Large-Scale Dataset: Cityscapes → BDD100K. “SF” refers to the source-free setting. “FRCNN” denotes variants based on the Faster R-CNN family; “DETR” denotes variants based on the DETR family. “Oracle” denotes an upper bound obtained by training on the target domain with ground-truth annotations.

Method	SF	Base	Person	Rider	Car	Truck	Bus	Train	Mcycle	Bicycle	mAP
Oracle	-	DETR	55.6	55.8	75.2	44.8	63.5	56.1	45.6	52.2	56.1
DA-Faster ( Chen et al., 2018 )	✗	FRCNN	29.2	40.4	43.4	19.7	38.3	28.5	23.7	32.7	32.0
SFA ( Wang et al., 2021 )	✗	DETR	46.5	48.6	62.6	25.1	46.2	29.4	28.3	44.0	41.3
AQT ( Huang et al., 2022 )	✗	DETR	49.3	52.3	64.4	27.7	53.7	46.5	36.0	46.4	47.1
AT ( Li et al., 2022b )	✗	FRCNN	43.7	54.1	62.3	31.9	54.4	49.3	35.2	47.9	47.4
MRT ( Zhao et al., 2023 )	✗	DETR	52.8	51.7	68.7	35.9	58.1	54.5	41.0	47.1	51.2

Table 2: Adaptation from Normal to Foggy Weather: Cityscapes → Foggy-Cityscapes.

Method	SF	Base	S2C	K2C
Oracle	-	DETR	79.9	79.9
DA-Faster ( Chen et al., 2018 )	✗	FRCNN	41.9	41.8
SFA ( Wang et al., 2021 )	✗	DETR	52.6	41.3
DATR ( Chen et al., 2025 )	✗	DETR	64.3	53.3
SFOD ( Li et al., 2021 )	✓	FRCNN	42.3	43.6
SFOD-M ( Li et al., 2021 )	✓	FRCNN	42.9	44.6

Table 3: Adaptation from Synthetic to Real Images: Sim10K → Cityscapes, Adaptation on cross-camera: KITTI → Cityscapes. The results are reported using the “car” class.

Depth	Number	Total Slot Numbers	mAP
2	2	4	50.9
2	4	16	52.9
2	5	25	53.2
2	6	36	52.6
2	8	64	52.5
2	10	100	51.2

Table 4: Hyperparameter Sensitivity Analysis. “Depth” denotes the hierarchy depth of HSA, and “Number” is the slot count per level. The total number of fine slots is

Number^{Depth}

Figure 6: Qualitative comparison of our method with the “Source Only” and TITAN. The red box is enlarged to highlight and compare the detection results more clearly. Our method captures fine-grained details more effectively, detecting distant objects that the other methods miss.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

\mathcal{L}_{\text{con}}=-\tfrac{1}{C}\sum_{c=1}^{C}\log\frac{\exp\!\bigl(\phi(P_{c},\bar{z}_{c})/\tau\bigr)}{\sum_{c^{\prime}=1}^{C}\exp\!\bigl(\phi(P_{c},\bar{z}_{c^{\prime}})/\tau\bigr)}

Appendix

Algorithm 3 Class-Guided Slot Contrast (CGSC)

Range	Method	mAP
$0.55{\rightarrow}0.40$	Cosine	53.2
$0.80{\rightarrow}0.40$	Cosine	45.7
$0.55{\rightarrow}0.20$	Cosine	51.2
$0.80{\rightarrow}0.20$	Cosine	47.0
$0.50$	Fixed	50.9
$0.40$	Fixed	48.9

Appendix

Table F.1: Comparison of threshold schedules. The cosine schedule with

0.55{\rightarrow}0.40

achieves the best result.

Figure G.1: Visualization of mask decomposition under different configurations. Top-left: Original image. Top-right: HSA ( $5\times 5$ slots). Bottom-left: SA (5 slots). Bottom-right: SA (25 slots). HSA achieves both stability and granularity, while vanilla SA either under-segments or collapses.

Method	mAP
SA (num_slots = 5)	50.6
SA (num_slots = 25)	50.2
HSA ( $5\times 5$ slots)	53.2

Appendix

Table G.1: Quantitative ablation on slot configurations (mAP@50). HSA consistently outperforms vanilla SA under both small and large slot settings.

Figure H.1: Additional visualizations of the slots.

Figure H.2: Visualization of the decomposition process

Figure I.1: Empirical validation of cosine-margin gain and reconstruction consistency

Setting	COCO pretraining for HSA	mAP
Source only	–	46.2
w/ HSA	$\times$	51.2
w/ HSA & CGSC (CGSA)	$\times$	52.7
w/ HSA	$\checkmark$	51.8
w/ HSA & CGSC (CGSA)	$\checkmark$	53.2

Appendix

Table J.1: Ablation Study of COCO pretraining for HSA.

Figure J.1: Visualization of distribution differences between COCO and all datasets used in our experiments.

Base Detector	Component	Parameters	Time	mAP
RT-DETR ( Zhao et al., 2024 )	–	43M	269ms	46.2
	+ HSA	63M	343ms	51.8
	$\Delta$ %	46.5%	27.5%	12.1%
FocalNet-DINO-DETR ( Yang et al., 2022 )	–	228M	780ms	43.8
	+ HSA	251M	809ms	55.0
	$\Delta$ %	10.1%	3.7%	25.6%

Appendix

Table K.1: Analysis of computational overhead. Results are measured on the Cityscapes → Foggy-Cityscapes setting.

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 619 papers

Contrastive Hypergraph Source-free Domain Adaptive Object Detection in Adverse Weather Conditions--Extended Version
Oct 17, 2026Jicheng Yuan, Tung Kieu, Manh Nguyen-Duc +2Object Detection Source-Free Domain Adaptation
Open Distributed Systems, Technical University of Berlin · Department of Computer Science, Aalborg University, Denmark · Fraunhofer Institute for Open Communication Systems, Berlin, Germany
ABRA: Teleporting Fine-Tuned Knowledge Across Domains for Open-Vocabulary Object Detection
Mar 12, 2026Mattia Bernardi, Chiara Cappellino, Matteo Mosconi +3Open-Vocabulary Object Detection Vision Foundation Model Adaptation
University of Modena and Reggio Emilia
DA-Mamba: Learning Domain-Aware State Space Model for Global-Local Alignment in Domain Adaptive Object Detection
Mar 19, 2026Haochen Li, Rui Zhang, Hantao Yao +5Object Detection Unsupervised Domain Adaptation
Intelligent Software Research Center, Institute of Software, CAS · University of Chinese Academy of Sciences · State Key Lab of Processors, Institute of Computing Technology, CAS +1

Oct 17, 2026J/K move · Enter open · S save

Contrastive Hypergraph Source-free Domain Adaptive Object Detection in Adverse Weather Conditions--Extended Version

Jicheng Yuan, Tung Kieu, Manh Nguyen-Duc, Manfred Hauswirth+1

Open Distributed Systems, Technical University of Berlin · Department of Computer Science, Aalborg University, Denmark · Fraunhofer Institute for Open Communication Systems, Berlin, Germany

Object Detection Source-Free Domain Adaptation Contrastive Learning

Open paper AskPDF arXiv

Adverse weather conditions such as fog, rain, snow, and low illumination introduce structured visibility degradations that severely impair object detectors trained on clear-weather data. When source data are inaccessible due to privacy or trans- mission constraints, source-free domain adaptation (SFDA) becomes necessary. However, existing SFDA methods based on pairwise contrastive learning struggle to capture high-order semantics and often suffer from agreement collapse under severe corruption. We propose REUNION, a hypergraph-guided SFDA framework for object detection in adverse weather conditions. REUNION constructs a contrastive hypergraph over target-domain object proposals, encoding high-order relations through intra-image context, weather-aware grouping, prototype-based semantic anchors, and uncertainty-aware connections between reliable and low-confidence instances. To effectively exploit these structured relations, we introduce a group-wise Hyper-InfoNCE objective that optimizes representations at the hyperedge level, enabling semantic information to propagate from confident proposals to corrupted or low-contrast instances. Experiments with diverse benchmarks demonstrate that REUNION effectively mitigates domain shifts and achieves state-of-the-art performance in SFDA settings.

Abstract

Object Detection

Explore similar work

Contrastive Hypergraph Source-free Domain Adaptive Object Detection in Adverse Weather Conditions--Extended Version

ABRA: Teleporting Fine-Tuned Knowledge Across Domains for Open-Vocabulary Object Detection

DA-Mamba: Learning Domain-Aware State Space Model for Global-Local Alignment in Domain Adaptive Object Detection