Category-aware Filtering for Open-set Semi-Supervised Object Detection

Semi-Supervised Object Detection (SSOD) has gained significant interest in recent years. However, the majority of existing methods are based on the close-set assumption. In this paper, we address a challenging but more practical scenario, Open-Set Semi-Supervised Object Detection (OSSOD), where out-of-distribution (OOD) samples are contained in unlabeled data. The previous approach employs an offline detector to eliminate OOD pseudo labels, which is intricate and time-consuming. And the limited availability of labeled data hampers the performance of the OOD detector. In contrast, we propose a simple yet effective approach that facilitates the detector’s performance in open scenarios. Specifically, we employ soft-distribution on all identified instances as our training objective to enhance the model’s feature extraction capability. And we further design a radius-aware matching approach to filter OOD samples according to categories with varying diversity in unlabeled data. Our category-aware design has been proven effective through extensive experiments, surpassing current state-of-the-art work.