Mass Estimation in Body Photography for Obesity Assessment Using Deep Learning and Linear Regression
摘要
This work presents a computer vision method for estimating the mass (weight) of people, based on the automatic interpretation of anthropometric attributes in low resolution, non-standardized almost full body photographs. Human body keypoints are obtained by deep learning, relationships between measurements are defined as features, and a regression architecture is trained for 405 images. When applying the model to another 100 unknown images, the correlation between actual and estimated mass measurements is 0.7216, with correct classification of obesity for 71% of cases.