Ambiguity-Resistant Semi-Supervised Learning for Dense Object Detection

With basic Semi-Supervised Object Detection (SSOD) techniques, one-stagedetectors generally obtain limited promotions compared with two-stage clusters.We experimentally find that the root lies in two kinds of ambiguities: (1)Selection ambiguity that selected pseudo labels are less accurate, sinceclassification scores cannot properly represent the localization quality. (2)Assignment ambiguity that samples are matched with improper labels inpseudo-label assignment, as the strategy is misguided by missed objects andinaccurate pseudo boxes. To tackle these problems, we propose aAmbiguity-Resistant Semi-supervised Learning (ARSL) for one-stage detectors.Specifically, to alleviate the selection ambiguity, Joint-Confidence Estimation(JCE) is proposed to jointly quantifies the classification and localizationquality of pseudo labels. As for the assignment ambiguity, Task-SeparationAssignment (TSA) is introduced to assign labels based on pixel-levelpredictions rather than unreliable pseudo boxes. It employs a"divide-and-conquer" strategy and separately exploits positives for theclassification and localization task, which is more robust to the assignmentambiguity. Comprehensive experiments demonstrate that ARSL effectivelymitigates the ambiguities and achieves state-of-the-art SSOD performance on MSCOCO and PASCAL VOC. Codes can be found athttps://github.com/PaddlePaddle/PaddleDetection.