Scale-Invariant Teaching for Semi-Supervised Object Detection

Recent Semi-Supervised Object Detection methods are mainly based on self-training, i.e., generating hard pseudo-labels by a teacher model on unlabeled data as supervisory signals. Although they achieved certain success, the massive False Negative samples and inferior localization precision lack consideration. Furthermore, the limited annotations in semi-supervised learning scale up the challenges: large variance of object sizes and class imbalance (i.e., the extreme ratio between background and object), hindering the performance of prior arts. We address these challenges by introducing a novel approach, Scale-Invariant Teaching (SIT), which is a simple yet effective end-to-end knowledge distillation framework robust to large object size variance and class imbalance. SIT has several appealing benefits compared to previous works. (1) SIT imposes a consistency regularization to reduce the prediction discrepancy between objects with different sizes. (2) The soft pseudo-label alleviates the noise problem from the False Negative samples and inferior localization precision. (3) A re-weighting strategy can implicitly screen the potential foreground regions from unlabeled data to reduce the effect of class imbalance. Extensive experiments show that SIT consistently outperforms the recent state-of-the-art methods and baseline on different datasets with significant margins. For example, it surpasses the supervised counterpart by more than 10 mAP when using 5% and 10% labeled data on MS-COCO.