User-Selected Object Data Augmentation for 6DOF CNN Localization

Automatic content placement is desired for augmented reality (AR) applications. Conventional approaches, such as marker deposition or three-dimensional modeling of all user-selected objects, are time-intensive. We herein propose a framework that is cost-efficient when the same content is superposed on the same multiple objects. The framework consists of steps: 1) scanning a user-selected object with an RGB-D SLAM, 2) cropping the object from the scanned images, and 3) training a convolutional neural network (CNN) on the cropped images and various backgrounds. Our CNN outputs the absolute position and tilt of a device, which assumes content placement in an AR application. An important aspect here is data augmentation. The framework creates images of a user-selected object taken from multiple viewpoints, without significant manual effort. The effectiveness of the proposed framework is also demonstrated.