FSDCNN: A few shot detection mechanism that preserves its supervised nature

Object detection has become better with the advent of deep convolution neutral networks. However, the challenge of training a fully supervised system when there is a small amount of training samples available still remains. Another issue with fully supervised systems is seen upon encountering novel classes. It is difficult to retrain the model as it is a time-consuming and tedious process. Inspired by a human’s ability to learn at a rapid rate, few-shot learning models have seen rapid development. In contrast to fully supervised systems, these learn from just a few samples. We propose a few-shot detection model, FSDCNN, based on a two-stage detector, that optimizes both the region proposal network and the object detector with the help of few-shot learning. FSDCNN performs similar to other models when only 1 or 3 new samples are seen but outperforms them when 5 or 10 samples of the new classes are seen, and it preserves the fully supervised nature of the base two stage detector.