Support-Query Mutual Promotion and Classification Correction Network for Few-Shot Object Detection

Recently, finetuning-based methods have shown great performance in few-shot object detection. These methods employ pure convolutional structures for feature extraction, and then finetune the last layers of the base detector on novel classes. However, most of them have the following two issues: 1) the feature extraction processes of support and query sets lack mutual promotion, and 2) the features extracted by pure convolutional structures have low discriminability for confused classes, which could lead to false positive results. To overcome these two issues, we propose a support-query mutual promotion and classification correction network (SMPCCNet). First, we build a support-query mutual promotion module. In this module, we perform class attention operations with query features to produce enhanced support features, and use the dense kernels generated by the enhanced support features to obtain support-relevant query features, which achieves the mutual promotion of support and query features. In addition, we design a classification correction branch with a hybrid attention module. The hybrid attention module uses spatial and channel attentions to generate features that focus on the positions and types of detected objects, respectively, which can better discriminate confused classes and correct false positive results. Extensive experiments demonstrate the superiority of SMPCCNet over state-of-the-art methods.