Attention guided deep features for accurate body mass index estimation

• An end-to-end deep learning framework is proposed to obtain BMI from 2D body images. • Combine attention mechanism with the deep network to enhance the performance. • Extensive experiments prove the proposed network outperforms the SOAT approaches. Body Mass Index (BMI) has been widely used as an indicator to evaluate the health condition of individuals, classifying a person as underweight, normal weight, overweight, or obese. Recently, several methods have been proposed to obtain BMI values based on the visual information, e.g., face images or 3D body images. These methods by extrapolating anthropometric features from face images or 3D body images are advanced in BMI estimation accuracy, however, they suffer from the difficulties of obtaining the required data due to the privacy issue or the 3D camera limitations. Moreover, these methods are hard to achieve satisfactory performance when they are directly applied to 2D body images. To tackle these problems, we propose to estimate BMI results from 2D body images by an end-to-end Convolutional Neural Network (CNN) with attention guidance. The proposed method is evaluated on our collected dataset. Extensive experiments confirm that the proposed framework outperforms state-of-the-art approaches in most cases.