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

Mattia Bernardi; Chiara Cappellino; Matteo Mosconi; Enver Sangineto; Angelo Porrello; Simone Calderara

cs.CVMar 12, 2026

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

Authors: Mattia Bernardi, Chiara Cappellino, Matteo Mosconi, Enver Sangineto, Angelo Porrello, Simone Calderara

Organizations: University of Modena and Reggio Emilia

Abstract

Although recent Open-Vocabulary Object Detection architectures, such as Grounding DINO, demonstrate strong zero-shot capabilities, their performance degrades significantly under domain shifts. Moreover, many domains of practical interest, such as nighttime or foggy scenes, lack large annotated datasets, preventing direct fine-tuning. In this paper, we introduce Aligned Basis Relocation for Adaptation(ABRA), a method that transfers class-specific detection knowledge from a labeled source domain to a target domain where no training images containing these classes are accessible. ABRA formulates this adaptation as a geometric transport problem in the weight space of a pretrained detector, aligning source and target domain experts to transport class-specific knowledge. Extensive experiments across challenging domain shifts demonstrate that ABRA successfully teleports class-level specialization under multiple adverse conditions. Our code will be made public upon acceptance.

Figures & tables

Figure 1: Goal. Transfer class-specific knowledge from the source domain to unavailable target classes without access to target-domain data.

Figure 2: Overview of the transfer pipeline. Starting from domain experts obtained through Objectification, class-specific residuals learned via SVFT in the source domain are analytically transported to unseen target domains.

Figure 3: Objectification. Each domain is fine-tuned using a unified “object” label to capture domain appearance.

	Bus		Motor		Rider		Train		Truck		Average
Method	mAP	AP 50	mAP	AP 50	mAP	AP 50	mAP	AP 50	mAP	AP 50	mAP	AP 50
Fine-tuning	58.75	73.50	31.22	55.01	43.87	68.06	32.95	58.75	40.03	57.08	41.36	62.48
Zero shot	48.63	61.10	23.20	41.91	18.96	31.70	16.31	44.02	31.20	41.88	27.66	44.12
Source	54.77	66.93	29.62	50.96	40.25	61.99	29.40	55.30	37.23	51.54	38.25	57.34
Task Analogy [ 11 ]	41.14	48.21	10.24	20.32	9.35	16.75	10.10	24.00	19.77	24.70	18.12	26.79
Param $\Delta$ [ 3 ]	50.23	60.41	20.59	40.59	18.53	31.44	21.70	49.42	30.42	40.26	28.29	44.42

Table 1: Cityscapes

\rightarrow

Foggy Cityscapes. Comparison of mAP and AP 50 on unavailable categories. ABRA outperforms all baselines. Best in bold .

	Day Foggy		Dusk Rainy		Night Clear		Night Rainy		Average
Method	mAP	AP 50	mAP	AP 50	mAP	AP 50	mAP	AP 50	mAP	AP 50
Fine-tuning	36.37	57.83	26.77	50.65	36.86	69.40	16.81	29.85	29.20	51.93
Zero shot	26.36	41.10	19.55	34.93	27.50	49.63	9.19	13.60	20.65	34.82
Source	31.75	51.41	27.63	50.34	36.38	66.70	15.28	27.52	27.76	48.99
Task Analogy [ 11 ]	26.49	41.27	18.97	34.69	27.38	49.44	9.61	14.27	20.61	34.92
Param $\Delta$ [ 3 ]	17.68	26.68	5.07	8.09	4.86	8.01	7.86	12.60	8.87	13.85

Table 2: SDGOD. Comparison of mAP and AP 50 across challenging condition splits. ABRA achieves the highest overall average among all baselines. Best in bold .

Figure 4: Effect of specialization on Cityscapes. Dedicating one expert per class (ABRA) yields higher AP 50 than using a single unified expert (Merge).

Panel of Figure 5 — Figure 5: Few-shot transfer evaluation. Cityscapes comparison (mAP) between the standard pretrained backbone ( $\theta_{0}$ ) and ABRA across varying numbers of shots.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Panel of Table 5 — Table 5: Top to bottom: bus, rider, train. Left to right: source (baseline), ABRA, GT

Open-Vocabulary Object Detection

Open-vocabulary object detection lets models locate and label objects using text categories beyond their fixed training labels.

See all 87 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
PISA: A Pseudo-Individual Source-Domain Feature Adaptation Framework for Test-Time Open-Vocabulary Object Detection
Aug 14, 2026Ziyan He, Xiongtai Yang, Tao WangOpen-Vocabulary Object Detection Test-Time Adaptation of VLMs
Sichuan University
The Detector Teaches Itself: Lightweight Self-Supervised Adaptation for Open-Vocabulary Object Detection
May 5, 2026Yazhe Wan, Changjae OhOpen-Vocabulary Object Detection VLM Adaptation
Queen Mary University of London · Queen Mary University of London, London, UK

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

Open-Vocabulary Object Detection

Explore similar work

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

PISA: A Pseudo-Individual Source-Domain Feature Adaptation Framework for Test-Time Open-Vocabulary Object Detection

The Detector Teaches Itself: Lightweight Self-Supervised Adaptation for Open-Vocabulary Object Detection