Towards Adaptive Open-Set Object Detection via Category-Level Collaboration Knowledge Mining

Yuqi Ji; Junjie Ke; Lihuo He; Lizhi Wang; Xinbo Gao

cs.CVApr 13, 2026

Towards Adaptive Open-Set Object Detection via Category-Level Collaboration Knowledge Mining

Authors: Yuqi Ji, Junjie Ke, Lihuo He, Lizhi Wang, Xinbo Gao

Organizations: School of Electronic Engineering, Xidian University, Xi’an 710071, China, and also with the Interdisciplinary Institute of Artificial Intelligence, Faculty of Information X, Xidian University, Xi’an, Shaanxi 710126, China. · School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China

Abstract

Existing object detectors often struggle to generalize across domains while adapting to emerging novel categories. Adaptive open-set object detection (AOOD) addresses this challenge by training on base categories in the source domain and adapting to both base and novel categories in the target domain without target annotations. However, current AOOD methods remain limited by weak cross-domain representations, ambiguity among novel categories, and source-domain feature bias. To address these issues, we propose a category-level collaboration knowledge mining strategy that exploits both inter-class and intra-class relationships across domains. Specifically, we construct a clustering-based memory bank to encode class prototypes, auxiliary features, and intra-class disparity information, and iteratively update it via unsupervised clustering to enhance category-level knowledge representation. We further design a base-to-novel selection metric to discover source-domain features related to novel categories and use them to initialize novel-category classifiers. In addition, an adaptive feature assignment strategy transfers the learned category-level knowledge to the target domain and asynchronously updates the memory bank to alleviate source-domain bias. Extensive experiments on multiple benchmarks show that our method consistently surpasses state-of-the-art AOOD methods by 1.1-5.5 mAP.

Figures & tables

Fig. 1: Illustration of (a) existing DAOD task, (b) OSOD task and (c) AOOD task. Foggy weather corresponds to the target domain, whereas clear weather corresponds to the source domain.

Fig. 2: Conceptual Visualization of the gap between mean features of base categories (novel category prediction) and real novel category features.

Fig. 3: An overview of the proposed method. In each mini-batch, images from the source domain (with ground truth annotations) and the target domain are used as input. A shared backbone and shared detector, based on Deformable DETR (DDETR), extract image-level and object query features. Object query features from the source domain are matched with ground truth to identify features corresponding to base categories. These matched features are then used to construct a clustering-based memory bank. Class prototype features guide the base-to-novel selection metric to identify novel category features, while class auxiliary features support adaptive feature assignment by generating high-quality pseudo-labels for object query features in the target domain.

Fig. 4: Update procedure of the clustering-based memory bank (CMB). (a) For base category, prototype features and auxiliary features are updated via K-means clustering. (b) For novel category, prototype features and intra-class disparity features are derived from base class statistics.

Fig. 5: Illustration of the feature distributions for base and novel categories in (a) SOMA and (b) the proposed CCKM, respectively. The shaded regions represent background areas. The purple star denotes the novel class prototype feature, while the green and blue stars indicate two distinct base class prototype features. Compared with SOMA, CCKM exhibits reduced overlap between base and novel feature distributions.

Method	Setting	Num. novel categories: 3				Num. novel categories: 4				Num. novel categories: 5
Method	Setting	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$
DDETR [ 23 ]	het-sem	47.52	0.00	0.341	459	45.24	0.00	0.506	1028	42.38	0.00	0.659	1968
PROSER [ 62 ]		46.92	1.80	0.271	218	44.19	2.02	0.415	531	41.99	2.00	0.584	1127
OpenDet [ 17 ]		47.04	1.92	0.269	221	45.71	1.89	0.499	511	42.09	1.70	0.579	922
OW-DETR [ 46 ]		43.31	1.84	0.432	192	42.52	2.10	0.619	451	39.92	1.98	0.684	814
SOMA [ 22 ]		50.87	3.78	0.268	139	48.06	4.41	0.412	340	45.55	4.08	0.526	649

TABLE I: Comparing with state-of-the-art validation results on Cityscapes

\rightarrow

Foggy Cityscapes. The top 2 results are shown in red , green .

Method	Setting	Num. novel categories: 3				Num. novel categories: 4				Num. novel categories: 5
Method	Setting	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$
DDETR [ 23 ]	het-sem	13.48	0.00	0.153	1448	13.49	0.00	0.164	1604	13.52	0.00	0.227	2378
PROSER [ 62 ]		13.32	1.53	0.148	910	13.35	1.48	0.163	1032	13.37	1.60	0.218	1466
OpenDet [ 17 ]		13.70	1.20	0.135	836	13.71	1.17	0.150	992	13.75	1.27	0.209	1244
OW-DETR [ 46 ]		13.15	1.27	0.129	792	13.15	1.27	0.157	908	13.50	1.30	0.201	1168
SOMA [ 22 ]		14.11	1.86	0.127	614	14.10	1.90	0.145	732	14.13	2.01	0.197	1074

TABLE II: Comparing with state-of-the-art validation results on Cityscapes

\rightarrow

BDD100k. The top 2 results are shown in red , green .

Method	Num.	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$
DDETR [ 23 ]	6	19.78	0.00	8.95	6347
PROSER [ 62 ]		18.23	32.37	9.87	5853
OpenDet [ 17 ]		20.57	41.15	8.93	4295
OW-DETR [ 46 ]		20.31	35.48	10.26	5184
SOMA [ 22 ]		21.70	43.15	7.32	4278
CCKM		23.70	36.72	6.77	3496

TABLE III: Comparing with state-of-the-art validation results on Pascal VOC

\rightarrow

Clipart.(num. indicates the number of novel classes.) The top 2 results are shown in red , green .

Fig. 6: The boxplot for Pascal VOC $\rightarrow$ CLipart, where the blue dots $\bullet$ and the red stars $\star$ indicate the performance of the baseline and CCKM, respectively.

BNSM	CMB	AFA	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$
Baseline (SOMA)			45.55	4.08	0.526	649
✔			46.92	3.15	0.511	813
	✔		45.61	4.34	0.524	579
		✔	46.47	3.35	0.641	634
✔	✔		47.15	4.26	0.497	730
	✔	✔	46.38	3.78	0.601	525

TABLE IV: Ablation study on Cityscapes

\rightarrow

Foggy Cityscapes under het-sem setting (5 novel classes). The best results are highlighted in bold.

SCM	TCM	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$
✔		46.65	2.55	0.574	974
	✔	46.36	3.31	0.568	824
✔	✔	46.92	3.15	0.511	813

TABLE V: Comparison with connection matrix (SCM :

U_{s}

and TCM :

U_{t}

) on Cityscapes

\rightarrow

Foggy Cityscapes het-sem setting (5 novel classes). The best results are highlighted in bold.

Constraint	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$
Cosine	44.83	3.09	0.544	834
ProtoBall	46.92	3.15	0.511	813

TABLE VI: Ablation Study on Prototype Modeling Strategies on Cityscapes

\rightarrow

Foggy Cityscapes under het-sem setting (5 novel classes). Cosine and ProtoBall denote cosine distance and ProtoBall distance, respectively. The best results are highlighted in bold.

$\beta$	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$
$1e-4$	45.42	3.26	0.485	492
$1e-3$	46.62	3.44	0.510	539
$1e-2$	47.79	4.16	0.494	500
$5e-2$	47.71	4.09	0.498	512

TABLE VII: Sensitivity analysis of the momentum parameter

\beta

on Cityscapes

\rightarrow

Foggy Cityscapes under het-sem setting (5 novel classes). The best results are highlighted in bold.

$K$	mAP $\uparrow$	AR $\uparrow$	WI $\downarrow$	AOSE $\downarrow$
$3$	48.13	3.91	0.560	631
$5$	47.79	4.16	0.494	500
$7$	47.54	4.12	0.538	528

TABLE VIII: Sensitivity analysis of the hyperparameter

K

on Cityscapes

\rightarrow

Foggy Cityscapes under het-sem setting (5 novel classes). The best results are highlighted in bold.

Fig. 7: Sensitivity analysis of the hyperparameter $\gamma$ for Cityscapes $\rightarrow$ Foggy Cityscapes (het-sem). The plot displays the variation curves of $mAP$ (top left), $AR$ (top right), $WI$ (bottom left), and $AOSE$ (bottom right) as the $\gamma$ value changes.

Fig. 8: t-SNE visualization of object query features on Cityscapes → Foggy Cityscapes under the freq-dec setting with 3 novel classes. (a) cosine distance; (b) ProtoBall distance for measuring unmatched object queries relative to base class prototype features.

Panel of Figure 17 — Fig. 9: Visualization of detection results in Cityscapes [ 36 ] $\rightarrow$ Foggy Cityscapes [ 37 ] dataset. The first column shows the groundtruth; the second and third columns visualize the detection results of SOMA [ 22 ] and CCKM, respectively. In the last column, the closeups of false positive (FP) and false negative (FN) errors are zoomed in red and yellow boxes. The detection errors of SOMA are described in the captions below each row.

Object Detection

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

See all 624 papers

ExDet: Open-Domain Open-Vocabulary Detection with Cross-modal Extrapolation and Rectification
Jun 8, 2026Yupeng Zhang, Yuzhong Feng, Ruize Han +3Open-Vocabulary Object Detection Object Detection
College of Intelligence and Computing, Tianjin University, Tianjin, China · Faculty of Computer Science and Artificial Intelligence, Shenzhen University of Advanced Technology, Shenzhen, China · School of Artificial Intelligence, Nanchang University, Jiangxi, China
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
Parameter-Efficient Semantic Augmentation for Enhancing Open-Vocabulary Object Detection
Apr 6, 2026Weihao Cao, Runqi Wang, Xiaoyue Duan +3Open-Vocabulary Object Detection Vision Foundation Model Adaptation
State Key Laboratory of Advanced Rail Autonomous Operation, Beijing Jiaotong University · Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, Beijing Jiaotong University · School of Computer Science & Technology, Beijing Jiaotong University +1

Jun 8, 2026J/K move · Enter open · S save

ExDet: Open-Domain Open-Vocabulary Detection with Cross-modal Extrapolation and Rectification

Yupeng Zhang, Yuzhong Feng, Ruize Han, Zhiwei Chen+2

College of Intelligence and Computing, Tianjin University, Tianjin, China · Faculty of Computer Science and Artificial Intelligence, Shenzhen University of Advanced Technology, Shenzhen, China · School of Artificial Intelligence, Nanchang University, Jiangxi, China

Open-Vocabulary Object Detection Object Detection Multimodal Object Detection

Open paper AskPDF arXiv

Open-domain open-vocabulary detection (ODOVD) requires detectors to generalize to both novel categories and unseen domains, making it more challenging than open-vocabulary detection. Existing methods typically train open-vocabulary detectors together with domain generalization modules from scratch, leading to high training cost. we propose ExDet, a lightweight category-domain collaborative generalization framework for ODOVD that enhances the cross-category and cross-domain generalization of existing detectors. ExDet consists of Text-Guided Extrapolation (TGE), a lightweight Detector-Compatible Rectification (DCR) module, and ExRPN. Specifically, TGE exploits the DeltaSpace property of vision-language models (VLMs) to infer category- and domain-aware proxy visual prototypes from text. DCR is learned from the TGE-generated prototypes in a detector training-free and real-data-free manner, and is inserted after the classification head at inference to rectify representations toward a detector-compatible source-domain visual distribution, thereby enhancing classification for targets from novel categories and unseen domains. ExRPN recalibrates proposal scores by combining semantic similarity with RPN confidence, improving recall for novel and domain-shifted objects while providing better support for subsequent classification and DCR. ExDet achieves SOTA performance on OD-LVIS, OV-LVIS, Objects365, and MSOSB.

Abstract

Object Detection

Explore similar work

ExDet: Open-Domain Open-Vocabulary Detection with Cross-modal Extrapolation and Rectification

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

Parameter-Efficient Semantic Augmentation for Enhancing Open-Vocabulary Object Detection