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

A new lightweight framework called ExDet aims to improve how object detectors handle both novel categories and unfamiliar visual domains without the high training costs of existing methods, according to a preprint posted to arXiv on June 8, 2026 [1]. The framework addresses open-domain open-vocabulary detection (ODOVD), a task that requires models to identify objects from categories they were not explicitly trained on, even when the images come from visual domains the model has not seen before [1][2]. Current approaches often train open-vocabulary detectors alongside domain generalization modules from scratch, a process the authors describe as costly [2]. ExDet instead functions as a bolt-on enhancement for existing detectors, improving their cross-category and cross-domain generalization [1][2]. The system is built from three components: Text-Guided Extrapolation (TGE), Detector-Compatible Rectification (DCR), and ExRPN [1][2]. TGE leverages a property of vision-language models called DeltaSpace to generate category- and domain-aware proxy visual prototypes directly from text descriptions [2]. DCR is then trained on these synthetic prototypes without requiring any real image data or retraining of the underlying detector [2]. At inference time, DCR is inserted after the classification head to shift the model's internal representations toward a source-domain visual distribution, which the authors state improves classification accuracy for novel categories and unseen domains [2]. The third module, ExRPN, recalibrates region proposal scores by fusing semantic similarity with the standard region proposal network confidence [2]. This adjustment is designed to increase recall for objects that are either from novel categories or appear in domain-shifted scenes, while also supplying better inputs for the downstream classification and rectification stages [2]. The preprint reports that ExDet achieves state-of-the-art performance on four benchmark datasets: OD-LVIS, OV-LVIS, Objects365, and MSOSB [1][2]. The paper was submitted to the Computer Vision and Pattern Recognition section of arXiv, an open-access repository that hosts preprints across physics, mathematics, computer science, and related fields [6]. As of late 2024, arXiv was receiving approximately 24,000 new submissions per month and had surpassed two million total articles at the end of 2021 [6].