Mutually Distilled Sparse RCNN for Few-Shot Object Detection

Few-shot object detection(FSOD), which conducted to detect novel object based on massive base class samples and few novel classes samples, has gained extensive research interest from academic and industry. Major existing FSOD approaches basically use Faster RCNN(FRCNN) as basic framework. However, the RPN of the FRCNN architecture generates redundant anchor frames, which leads to slow training and consume massive computing resources. In this paper, we choose Sparse RCNN which has a fixed number of anchor frames and good performances in object detection the basic framework. Traditional FSOD training methods customarily use the method of freezing backbone or deleting the head classification branches, which will lead to overfitting of novel class and perform badly on base classes. To solve this problem, we introduced the mutual distillation layer and IOU Mask module in the head and loss of Sparse RCNN respectively. The mutual distillation layer is a multihead structure, which can distill the base class features when training new class samples, and distill the new class features when training base class samples. Experiments on multiple benchmarks show that our framework is significantly superior to other existing methods, and has faster detection speed.