Dual-Path Learning Toward Open-Vocabulary Object Detection in Remote Sensing Images

Abstract

While the landscape of closed-set oriented object detection has been revolutionized in recent years, extending this success to the instances of unseen categories, the core challenge of open-vocabulary object detection (OVD), remains a formidable hurdle. Prevailing methods often resort to a dedicated student detector to channel the cross-modal knowledge from pretrained vision-language models (VLMs), thereby enabling open-vocabulary generalization. However, these approaches, which primarily manifest as either knowledge distillation or pseudo-labeling frameworks, suffer from a fundamental tradeoff: biased region proposal network (RPN) yields poor recall for target objects, while shared detection head induces catastrophic forgetting of base classes. To overcome these limitations, we propose DPNet, a dedicated end-to-end dual-path network for open-vocabulary oriented object detection in remote sensing scenarios. First, we introduce a parameter-efficient multiscale adapter (MSA) that activates hierarchical features from a frozen vision-language encoder, enabling a category-agnostic network to generate high-quality proposals for target objects with varying scales. Second, a decoupled detection pipeline is formulated for base and target classes. This divide-and-conquer fashion alleviates the inherent optimization conflict, preserving base-class recognition fidelity while effectively learning target concepts. By circumventing the need for a student detector, DPNet offers a more streamlined and efficient architecture. Extensive experiments on two standard remote sensing benchmarks—DIOR and DOTA—validate that our solution achieves competitive performance, significantly outperforming pioneering approaches in both base-class precision and novel-class discovery under zero-shot settings. The source code will be available at https://github.com/yanqingyao1994/DPNet