Few-Shot Object Detection using Global Attention and Support Attention

General object detection methods in computer vision task typically require a lot of bounding box labeled data. In addition, after training on the training data of the specified category, only the existing categories in the training can be detected. In order to detect the new category, sufficient training data of the new category must be collected and the new category can only be detected after the network is trained again. In this paper, we design a few-shot object detection network, aiming to achieve object detection against unseen categories by using a small number of annotated sample data as support sets. Our network is based on Faster R-CNN, and using a Fixed-size Global Attention module on the features generated by the backbone so that the whole network has different attention weights for different positions of the image. A Support-Set Attention module is proposed which enables the network to focus the detection target on the categories provided by the support set. Finally, the relation detection module uses the similarity between the few-shot support set and the query set to give the matching score and the location regression. After training on the base class, our network can be directly used on N-way K-shot detection without fine-tuning. In our experiments, our network achieved equal or better results than others.