Benefiting From OOD Samples in Open-Set Semi-Supervised Object Detection

Open-set semi-supervised object detection (OSSOD) is an emerging research area that relaxes the assumption of closed-set in semi-supervised object detection (SSOD), allowing unlabeled data to contain both in-distribution (ID) and out-of-distribution (OOD) samples. In the face of OSSOD, existing approaches expect to completely filter out OOD samples from unlabeled data for conventional semi-supervised learning. Different from it, we find that the effective use of OOD samples in unlabeled data can also facilitate feature learning and thus boost the detection performance of ID categories under open set conditions. Specifically, we perform refined instance-level consistency regularization (ICR) on all the detected instances (including the OOD ones) on unlabeled images, while introducing a novel OOD-aware contrastive learning (OCL) that clusters ID objects and pushes away OOD samples in the intra-class feature space, enhancing the compactness of ID features while emphasizing the discrimination between ID and OOD objects. Based on the distinguished features, we further design a prototype-based multimetric adaptive matching (MAM) approach, which identifies ID/OOD samples by adaptively measuring multiscale feature similarity between samples and class prototypes according to the specific category. With the help of this approach, more reliable ID/OOD samples can be mined from unlabeled data for contrastive learning, while the consistency regularization (CR) can be weighted according to the matching scores, correcting the biasing effects of OOD objects. Extensive experiments demonstrate that our design can effectively utilize OOD samples to optimize feature learning while avoiding their detrimental effects in semi-supervised learning, which ultimately greatly improves the detection capability of the model in open-set scenarios by a large margin over current state-of-the-art works.