Few-shot Small Object Detection in Predictable Background

In certain application domains, there is a need to detect objects from simple and uniform backgrounds, such as detecting birds in the sky or foreign objects on the ground. The difficulty of these detection tasks lies in the scarcity of targets and small size. However, a common characteristic of these detection tasks is that background information is usually predictable. Based on the predictability of the background information, this study investigates the use of classical learning and deep learning methods from different feature perspectives to accomplish few-shot and small object detection. The results demonstrates that the selected classic methods did satisfaction in detecting few-shot small objects in simple and predictable background but are sensitive to noise or tiny changes from backgrounds, while chosen deep learning are relatively resistant to backgrounds variations and show better ability in generalization.