[CVPR 2022] semi supervised object detection: dense learning based semi supervised object detection

brief introduction ：

Semi supervised target detection （SSOD） Designed to take advantage of large amounts of unlabeled data , Promote the training and deployment of target detectors . Although various methods based on self training and consistent regularization have been proposed SSOD Method , But most of them are anchor frame based detectors , However, it ignores the fact that in many practical applications, there are higher requirements for anchor free frame detectors .

In this paper , The author intends to fill this gap , And a method based on Dense Learning No anchor for SSOD Algorithm .

say concretely , The author introduces several new technologies , Include ：1、 Adaptive filtering （Adaptive Filtering） The strategy assigns multiple layers of accurate dense pixel level pseudo tags ;2、 Gather teachers （Aggregated Teacher） Used to generate stable and accurate pseudo tags ;3、