Enhancing Open-Set Object Detection via Uncertainty-Boxes Identification

Open-set object detection is a challenging task in computer vision, which aims to detect known object categories while simultaneously identifying unknown objects. Inspired by how humans naturally distinguish unseen objects by comparing their similarities and dissimilarities with known objects, we propose a novel network called UNBDet (Uncertainty-Boxes Detection) for enhancing open-set object detection. Our approach, UNBDet, utilizes a Pseudo Proposal Advisor (PPA) to generate a wide range of unknown object candidates to improve learning and make the distribution of pseudo unknowns more consistent with the actual unknown object distribution. Furthermore, we employ an Unknown Probability Estimator (UPE) and Uncertainty-NMS modules to reason the predicted overlapping and uncertainty-boxes from multiple known classes, thus enabling easy identification of unknown objects with high uncertainty. Our experimental results demonstrate that UNBDet significantly outperforms state-of-the-art models in open-set object detection.