Rethinking Object Detection: A Unified Framework for Detection Refinement

Recent advances in object detection have primarily focused on designing sophisticated neural networks to improve detection accuracy. Although these models have achieved remarkable performance, their outputs still contain a substantial number of false positives. Existing approaches primarily focus on improving the prediction capabilities of detection networks, while overlooking the importance of refining their outputs for practical applications. In this paper, we rethink object detection and introduce a unified refinement framework that can be seamlessly integrated into existing detection models. Specifically, our framework employs a group-aware refinement strategy to substantially reduce false positives while retaining valid detections. Furthermore, we establish theoretical guarantees for the proposed method. Extensive experiments demonstrate that our framework effectively reduces false positives while improving average precision.