3D-FREDI

Abstract

3D face reconstruction has been a challenging task for the last two decades in the field Computer Vision. Standard approach to 3D facial reconstruction relies on the availability of multiple facial images, but this approach presents itself with a number of challenges. Such as establishing desne correspondence between different facial expressions,poses and camera angles. Illumination problems also introduce errors in the reconstructed 3d model. In 3D-FREDI (3D Facial reconstruction and emotional detection with interpretation) we propose a novel solution to this problem by reconstructing human face model using a monocular image. We address many of the limitations of the standard approach by training a Deep Convolutional Generative Adversarial Network a dataset of 2D images and their corresponding depth images. Our DCGAN bypasses the problems such as dense correspondence between different facial poses, expressions. The generated 3D models are then fed into a regressor which performs human emotion classification on the generated depth image. The first portion is concerned with the conversion of a simple 2D image into a realistic 3D model. For brevity’s sake, we will not elaborate on our approach here, preferring to do so in the appropriate sections of this document. Suffice to say, we utilized Deep Convolutional Generative Adversarial Network (DCGAN) in order to generate a 3D model from a 2D image, subsequently using a convolutional neural network (CNN) to perform emotion classification on the former. To succinctly state our endeavors, we attempted to generate a 3D representation of a human face and subsequently classify the emotional state of the individual from the 3D models.