Temporal Self-Ensembling Teacher for Semi-Supervised Object Detection

This paper focuses on the semi-supervised object detection (SSOD) which makes good use of unlabeled data to boost performance. We face following obstacles when adapting knowledge distillation (KD) framework in SSOD. (1) The teacher model serves a dual role as and student, such that predictions images may limit upper bound student. (2) imbalance issue caused by large quantity consistent between student hinders an efficient transfer them. To mitigate these issues, we propose novel SSOD called Temporal Self-Ensembling Teacher (TSET). Our ensembles its temporal for under stochastic perturbations. Then, our weights with those exponential moving average. These ensembling strategies ensure diversity, lead better images. In addition, adapt focal loss formulate consistency handling issue. Together thresholding method, automatically reweights inconsistent predictions, preserves difficult objects detect mAP reaches 80.73% 40.52% VOC2007 test set COCO2014 minival5k set, respectively, outperforms strong fully supervised detector 2.37% 1.49%, respectively. Furthermore, (80.73%) sets new state-of-the-art performance set.