Object Detection with Few-Shot Learning and Data Augmentation

Deep learning based object detection techniques require large amount of training data and long training time. However, the availability of large annotated image dataset is usually limited and expensive to generate. Moreover, training with large dataset is slow and thus not suitable for real-time applications such as video surveillance which need to learn to detect new object categories or adapt to environmental changes quickly. In this paper, a framework is proposed for object detection with little data samples using few-shot learning and data augmentation. The proposed framework adopts the meta-learning paradigm where a base detection network first learns to extract meta-features from the base classes with abundant training samples and then an adaptation network is trained to generate class-specific weights to adapt the meta-features to detecting novel object classes using only a few support samples. Moreover, data augmentation is applied on the support set to boost the detection performance further. The proposed framework shows superior few-shot object detection performance over state-of-the-art methods on benchmark dataset.