FEW-SHOT IMAGE RECOGNITION AND OBJECT DETECTION

Abstract

Image recognition and object detection are fundamental vision tasks which are important for many real-world applications, such as robot grasping, manipulation, navigation, etc. However, the successful deployment of deep learning systems on real-world applications still faces challenges in model generalization and learning efficiency. Motivated by this, we study the problem of few-shot learning and its boarder applications in object detection, which are important for many real-world industrial applications. Our goal is to design efficient data-efficient models that can learn to recognize and localize novel objects with limited supervision signals. We start from the task of few-shot detection, which aims at adapting a pre-trained object detector to additional localize novel-class instances. Then, we tackle a more challenging task named as incremental few-shot learning, i.e. novel classes come in a sequential manner instead of being provided as a whole, which requires to continually fine-tuning a model on small training samples.