A Multi-view Images Generation Method for Object Recognition

摘要

Recently convolutional neural network has achieved great success in the field of object recognition, but it is a hard work to get enough labeled data for training a neural network, especially for novel object instances. In this paper, we address this problem by generating synthetic images in simulation environment. We propose a method that generates a large amount multi-view synthetic images automatically to avoid manual collection and annotation. When applying our method to object recognition in real scenarios, the robot picks up the object first, then gets the object images using the same method of getting training images, which reduces the domain gap between real images and synthetic images. Experiments show that our method can recognize various objects with different poses efficiently.