Reconhecimento facial RGBD para análise de parentesco

Abstract

This work presents a new approach to kinship recognition based on Deep Learning applied to facial data of color images with depth information, i. e., RGBD. To work around the lack of an adequate 3D database containing kinship information, an online platform was provided where participants can submit videos captured by common smartphones cameras containing their face and those of their relatives. Then, the videos are processed to generate the 3D reconstruction of recorded faces, resulting in a standardized database coined Kin3D. It combines depth information from normalized 3D reconstructions with 2D images comprising RGBD data with unprecedent kinship information. Following previous works, image files are segmented into four categories according to their respective kinship relationship. For the classification, Convolutional Neural Networks (CNN) were used, as well as a Support Vector Machines (SVM) to obtain a baseline. The CNN was tested in a 2D kinship database previously consolidated in the scientific literature, known as KinFaceW-I and II, and in our Kin3D for comparison with related works. Another approach was used by bringing all first-degree relatives together at once and classifying them in a binary way. Results have shown that the addition of depth information improves the performance of the model, increasing the classification accuracy. As of the writing of this work, this is the first database containing depth information for kinship verification as well as the analysis of state-of-the-art techniques for obtaining the benchmark, providing performance as a starting point to further stimulate evaluations from the research community.