2024/12/13 by Haomiao Liu, Xu Hao, Liu, Haomiao +5 · 1 citation
Engineering · #Robotics and Automated Systems
paper · pdf · doi:10.48550/arxiv.2412.10176
Unknown Object Detection (UOD) aims to identify objects of unseen categories,\ndiffering from the traditional detection paradigm limited by the closed-world\nassumption. A key component of UOD is learning a generalized representation,\ni.e. objectness for both known and unknown categories to distinguish and\nlocalize objects from the background in a class-agnostic manner. However,\nprevious methods obtain supervision signals for learning objectness in\nisolation from either localization or classification information, leading to\npoor performance for UOD. To address this issue, we propose a transformer-based\nUOD framework, UN-DETR. Based on this, we craft Instance Presence Score (IPS)\nto represent the probability of an object's presence. For the purpose of\ninformation complementarity, IPS employs a strategy of joint supervised\nlearning, integrating attributes representing general objectness from the\npositional and the categorical latent space as supervision signals. To enhance\nIPS learning, we introduce a one-to-many assignment strategy to incorporate\nmore supervision. Then, we propose Unbiased Query Selection to provide premium\ninitial query vectors for the decoder. Additionally, we propose an IPS-guided\npost-process strategy to filter redundant boxes and correct classification\npredictions for known and unknown objects. Finally, we pretrain the entire\nUN-DETR in an unsupervised manner, in order to obtain objectness prior. Our\nUN-DETR is comprehensively evaluated on multiple UOD and known detection\nbenchmarks, demonstrating its effectiveness and achieving state-of-the-art\nperformance.\n