Double Head Predictor based Few-Shot Object Detection for Aerial Imagery
Many applications based on aerial imagery rely on accurate object detection, which requires a high number of annotated training data. However, the number of annotated training data is often limited. In this paper, we propose a novel few-shot detection method for aerial imagery that aims at detecting objects of unseen classes with only a few annotated examples. For this purpose, we extend the Two-Stage Fine-Tuning Approach (TFA), which achieves state-of-the-art results on common benchmark datasets. We pro-pose a novel annotation sampling and pre-processing strategy to yield a better exploitation of base class annotations and a more stable training. We further apply a modified fine-tuning scheme to reduce the number of missed detections. To prevent loss of knowledge learned during the base training, we introduce a novel double head predictor, yielding the best trade-off in detection accuracy between the novel and base classes. Our proposed Double Head Few-Shot Detection (DH-FSDet) method outperforms state-of-the-art baselines on publicly available aerial imagery datasets. Finally, ablation experiments are performed in or-der to get better insight how few-shot detection in aerial imagery is affected by the selection of base and novel classes. We provide the source code at https://github.com/Jonas-Meier/FrustratinglySimpleFsDet.