[ICLR 2021] semi supervised object detection: unbiased teacher for semi supervised object detection

Abstract ：

Semi-supervised learning , That is, the training network using both labeled and unlabeled data , Significant progress has been made recently .

However , The existing work mainly focuses on the task of image classification , The target detection is ignored .

So in this work , Semi supervised object detection is discussed again （SSOD）, And determined SSOD The problem of false mark deviation in . To solve this problem , The author introduces the unbiased teacher model （Unbiased Teacher）, This is a simple and effective method , Train the student model and the progressive teacher model in a mutually beneficial way .

In addition, it reduces the category balance loss of over fitted pseudo tags , The unbiased teacher model is in COCO and VOC The most advanced methods have been significantly improved on the data set .

Paper introduction ：

The availability of large-scale data sets and computing resources enables deep neural networks to achieve strong performance in a variety of tasks . However , Training these networks requires a large number of tagged instances , This data requires expensive annotation and acquisition .

As an alternative , Semi-supervised learning （SSL） Fang