Few-Shot Object Detection of Remote Sensing Images via Two-Stage Fine-Tuning

In recent years, there have been many developments in object detection on remote sensing images. However, those deep convolutional neural network (CNN) models always demand a great number of labeled samples, which leads to a significant decrease in performance on rare categories. Recently, the fine-tuning-based method of few-shot learning has drawn attention in the field of computer vision. In this letter, we proposed a multiscale few-shot object detection approach for remote sensing images, which is built upon faster region-CNN (R-CNN) architecture. First, we build the whole backbone of the detector with the involution operator, which enhances the classification ability of the features extractor. Then, our proposed detector learns multiscale features with the assistance of a path-aggregation module, which shortens the information transmission path by a bottom-up flow and uses semantic information of low-level features for localization. Eventually, we increase the shape bias in the detector in the training phase, which further improves the robustness and performance of the model. Experiments on two optical remote sensing datasets demonstrate that the proposed method outperforms the current few-shot detection models in the remote sensing field.