Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : June 02, 2018 - June 05, 2018
Recent years, Deep Learning is known as the method to achieve high recognition accuracy, and its application to robot technology has been actively performed. R-CNN (Regions with Convolutional Neural Networks) is well known as one of Deep Learning methods. However, a number of images and location labels are necessary for R-CNN. Although superimposing the object image randomly on the scene can solve this problem, images obtained by this method are weird ones. The precision of the classifier using weird images would be worse. In this article, an automatic learning image method which use AR marker and 3D model is proposed. The real object is scanned using a mobile phone to make a 3D model, and this model is virtually located using an augmented reality (AR) marker, which is put on the real environment naturally. Doing so, many of more contextual, i.e., natural, images including the object can be obtained. It was experimentally proved that the classifier which is learned by this method can search and detect objects in the real world with considerable high probability.
