| Resumo: | Intelligent Systems are designed to substitute the human component therefore they have a need to emulate a human's ability to quickly estimate biological traits of others, which is an integral part of social interactions. Age is one of the key characteristics used by marketing, entertainment and security tools. Existing age estimation systems can be easily fooled due to their reliance on human appearance based features, which can be easily manipulated. Over the years, while the complexity of models increased, the data fed to our systems was kept the same: a single 2D RGB image. This thesis addresses the current lack of studies made on the uses of 3D facial information ion the context of age estimation. This thesis encompasses a comprehensive study of how different 3D facial features can be used to improve current state of the art age estimation approaches using Deep Learning. Along with extensions to a baseline Convolutional Neural Network (CNN) model with a 2D image input, it is introduced a novel Multi-View CNN model which combines face descriptors from multiple perspectives within the model's architecture. Due to lack of 3D facial datasets aimed at age estimation, 2D age estimation datasets were synthetically augmented with landmark localization, 3DMM parametrization and 3D facial point cloud reconstruction. The last one was subsequently used to create a new synthetic dataset composed of renderings of each point cloud from different camera positions. A fully customizable data processing tool was introduced which supports image pre-processing, dataset splitting, image augmentation and synthetic feature extraction. Quantitative results show improvement of the 3D methods over traditional 2D although somewhat constrained by data quality.
|
|---|