Improving Few-Shot Object Detection through a Performance Analysis on Aerial and Natural Images

Object detection models require a large amount of annotated data during training, making their deployment for real-world tasks difficult. Few-Shot Object Detection (FSOD) aims to solve this shortcoming by training object detection models from limited data. However, existing methods mostly focus on natural images such as MS COCO and Pascal VOC datasets. In this paper, we study FSOD on aerial images. At first glance, performance seems to decrease compared to natural images with similar datasets (i.e. same number of images and classes). We perform an in-depth analysis to understand the performance discrepancies between natural and aerial images. In the light of this analysis, we propose several improvements to boost the detection quality on aerial images: new data augmentations for object detection, and new support cropping strategies. These modifications increase the mAP by approximately 5% on average.