Fine-tuning Faster-RCNN tailored to Feature Reweighting for Few-shot Object Detection

Few-shot object detection has drawn more attention in computer vision now. One acknowledged task setting is that train the network to detect both of the base classes with abundant images and novel classes with only a few. Under this scenario, two classic pipelines of few-shot object detection are developed. One is fine-tuning, which trains the detection network on images of base classes and fine-tunes the last layer on images of both base and novel classes. The other one is meta learning, in which one pioneering model utilizes a meta-learner to transform supporting images into reweighting vectors, which are used to reweight features of the query images obtained through the feature extractor. A typical meta learning method splits the training process into two phases: meta-training and meta-testing. Firstly in the meta-training phase, the model is trained on base classes, then on both base and novel classes. In this paper, we synthesize these two pipelines together. For the network structure, we tailor Faster-RCNN to the reweighting module; for training, we follow the meta-training procedure and fine-tune the reweighting module and only the last layer of Faster-RCNN during meta-testing. Experiments on NWPU VHR-10 images show that our method improves the mAP by about 10 ∼ 20 percentages than both of the reweighting and fine-tuning methods.