BMI estimation from facial images using residual regression model

Abstract

Body Mass Index (BMI) has the potential to disclose a variety of health and lifestyle concerns. Predicting BMI from facial images is an interesting but challenging problem in computer vision. Previous works focus mainly on feature extraction step of the whole BMI estimation process. Little attention has been paid to the regression module. In this paper, we propose a new architecture for the regression module which composes of multiple blocks. Each block has several sub-blocks composing of dense layer, batch-normalization, activation, dropout. In addition, we take advantage of the residual principle from ResNet by adding residual connections in the regression blocks. We integrate the proposed regression model just after the state-of-the-art feature extractor ResNet and train the network in an end-to-end manner. Extensive experiments on the VIP Attributes dataset show that thanks to the new residual regression model, the estimation error reduces up to 22% in comparison to the original method.