URSD: A uncertainty-resistant semi-supervised approach for object detection

With the development of mean teacher self-training, semi-supervised object detection (SSOD) has achieved significant progress. However, we experimentally find that further performance gains remain challenging due to two types of uncertainties: (1) sample mining uncertainty that discarding a large proportion of unlabeled samples based on manually setting confidence thresholding, ignoring valuable samples with lower confidence; (2) anchor assignment uncertainty that cannot resist noisy pseudo-bounding boxes based on static IoU-based strategy, undermining efficient learning on unlabeled data. To tackle these issues, we propose a novel uncertainty-resistant semi-supervised approach for object detection (URSD). Specifically, to alleviate the sample mining uncertainty, a confidence-aware adaptive weighting (CAW) mechanism is proposed to quantify the confidence of pseudo-labels as the weights for unlabeled samples mining. As for anchor assignment uncertainty, a noise-robust anchor assignment (NRA) is designed to comprehensively consider the noise of pseudo-boxes, pseudo-labels, and label assignment. Experimental results on MS COCO and PASCAL VOC datasets demonstrate the superior performance of the proposed URSD.