Abstract
Open-vocabulary detection (OVD) aims to detect and classify objects from an unrestricted set of categories, including those unseen during training. Existing open-vocabulary detectors often suffer from visual-textual misalignment and long-tailed category imbalance, leading to poor performance when handling objects described by complex, long-tailed textual queries. To overcome these challenges, we propose Multimodal Question Answering Detection (MQADet), a universal plug-and-play paradigm that enhances existing open-vocabulary detectors by leveraging the cross-modal reasoning capabilities of multimodal large language models (MLLMs). MQADet can be seamlessly integrated with pre-trained object detectors without requiring additional training or fine-tuning. Specifically, we design a novel three-stage Multimodal Question Answering (MQA) pipeline that guides MLLMs to accurately localize objects described by complex textual queries while refining the focus of existing detectors toward semantically relevant regions. To evaluate our approach, we construct a comprehensive benchmark across four challenging open-vocabulary datasets and integrate three state-of-the-art detectors as baselines. Extensive experiments demonstrate that MQADet consistently improves detection accuracy, particularly for unseen and linguistically complex categories, across diverse and challenging scenarios. To support further research, we will publicly release our code.