Virtual mirror by fusing multiple RGB-D cameras

Mirror is possibly the most common optical device in our everyday life. Rendering a virtual mirror using a joint camera-display system has a wide range of applications from cosmetics to medicine. Existing works focus primarily on simple modification of the mirror images of body parts and provide no or limited range of viewpoint dependent rendering. In this paper, we propose a framework for rendering mirror images from a virtual mirror based on 3D point clouds and color texture captured from a network of structured-light RGB-D cameras. We validate our models by comparing the results with a real mirror. Commodity structured-light cameras often have missing and erroneous depth data which directly affect the quality of the rendering. We address this problem via a novel probabilistic model that accurately separates foreground objects from background scene before correcting the erroneous depth data. We experimentally demonstrate that our depth correction algorithm outperforms other state-of-the-art techniques.