Body Mass Index (BMI) Estimation from Multi-view Facial Images

Accurate estimation of Body Mass Index (BMI) from facial images holds significant potential for applications in health and fitness monitoring. This study explores the efficacy of Deep Learning models in estimating BMI from facial images through two experiments: one utilizing solely frontal-view images and another combining both frontal and profile views (Multi-view). Performance is assessed using three Deep Learning models namely ResNet-50, MobileNet-V2, and DenseNet-121, with evaluation metrics including Mean Absolute Error (MAE), Mean Squared Error (MSE), and R-squared ($\mathrm{R}^{2}$). The models are trained and validated on a public IDOC dataset comprising 140,000 images. The experimental results indicate that incorporating frontal and profile views (multi-view) significantly enhances the accuracy. Among all, the ResNet-50 model achieves the lowest MAE (2.60) and MSE (12.60), and the highest $\mathrm{R}^{2}$ value (0.52) when multi-view images are used. Conversely, with only frontal views, MobileNet-V2 demonstrates the best performance in MAE (2.71) and nearly in MSE (13.73), while also maintaining a high $\mathrm{R}^{2}$ value (0.51). These results underscore the potential of Deep Learning models in improving BMI estimation from facial images, which could be pivotal for developing more precise health and fitness monitoring systems and medical applications.