Hand Segmentation for Augmented Reality

Tadej Vodopivec (2016) Hand Segmentation for Augmented Reality. MSc thesis.

Preview

PDF Download (20Mb)

Abstract

Occlusion detection is a very important part of augmented reality because it allows us to render convincing compositions of real and virtual objects. The hardest part of creating such composition is to detect when real objects lie between the user and the virtual object. Because hands are often in our field of view, it is important to accurately detect their position to determine which parts of the virtual objects should be visible. In this paper we describe a method for hand segmentation based on a convolutional neural network. With this method we were able to efficiently and accurately detect the area where the hands were directly visible in a set of first-person view images. The images ranged from outdoors to an office-like environment. We expect the method to make the biggest impact in the field of augmented reality, where the user wears glasses-based AR system and interacts with the world with his hands.

Actions (login required)