Few-shot object detection on aerial imagery via deep metric learning and knowledge inheritance

Object detection is crucial in aerial imagery analysis. Previous methods based on convolutional neural networks (CNNs) require large-scale labeled datasets for training to achieve significant success. However, the acquisition and manual annotation of such data is time-consuming and expensive. In this study, we present an original few-shot object detection (FSOD) method that focuses on detecting unseen objects in aerial imagery with limited labeled samples. Specifically, we revisited the multi-similarity network from deep metric learning and incorporated it into a faster region-CNN (R-CNN) architecture for FSOD, learning distinctive feature representations, and effectively improving the performance of unseen class samples. Furthermore, we preserved the knowledge learned from abundant base data by designing a knowledge inheritance module to ease the influence of catastrophic forgetting. We conducted experiments on two benchmark remote sensing image datasets, and the results demonstrated that the proposed methods could achieve a satisfactory performance for FSOD in aerial imagery.