Efficient Feature Enhancement for Few-Shot Object Detection
Lin Li, Lei Zhou, Shengbo Chen, Qingguo Xu · 2022 IEEE 6th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC ) · 2022
The neural network has exhibited strong performance in various machine learning tasks, but deep neural network model will deteriorate rapidly when the amount of data is insufficient. Few-shot object detection aims to detect unseen objects under little supervision scenarios. Attention RPN uses support images to enhance the query image, enabling RPN to generate support-relevant bounding boxes and to detect novel categories without fine-tuning. In this paper, We propose a more powerful approach to enhance query feature. Firstly, we explored approaches to generate convolutional kernels containing support information and selected the more useful ones. Then, as the Attention RPN does not take into account the correlation between channels, we added the channel attention associated with the support feature to the query feature to further improve the detection of the model. To verify the effectiveness of the method proposed in this paper, we conducted experiments on the benchmarks and showed good results.