多視点画像を用いたメタ学習によるOne-Shot学習

抄録

Object detection and recognition are essential for robot manipulation. This paper presents one-shot image identification via meta-learning from multi-view images. In contrast to general image recognition tasks, a robot deployed in the real world has access to limited data of unseen objects. Another difficulty of object recognition in robot manipulation is variation of appearance due to a move of a camera. We propose to leverage meta-learning method to initialize fast-adaptable parameters to a single image of a new object. We train a network using multi-view images in pre-training to learn initial parameters that is fast-adaptable to multi-view images of a new object. Simulation is conducted to compare accuracy of the proposed method with variants of it.