Abstract
A person's physical characteristics, such as weight and height, serve as important indicators of physical and mental health, daily lifestyle, and financial well-being. Body Mass Index (BMI) is a widely recognized measure that encodes both weight and height information. BMI is commonly used as a self-monitoring tool and has long-term implications for health outcomes, including the prediction of disease risk and estimation of longevity. Automatically estimating BMI from a single image of a person captured in unconstrained, real-world environments remains challenging due to variations in human pose, camera geometry, appearance, and background complexity.
This paper explores the performance of deep neural networks using both single-task and multi-task learning approaches. We employ multiple input modalities—RGB images, depth maps, pose affinity maps, and edge maps—to predict BMI, weight, and height from everyday images sourced from social networking platforms. Since no comprehensive full-body image dataset for BMI estimation is publicly available, we introduce a new dataset comprising 6,105 images with ground-truth labels for height, weight, and BMI. Our dataset is collected in the wild and includes individuals from diverse ethnicities, age groups, and genders. It features frontal, back, full-body, half-body, and side poses, as well as mirror selfies with varying backgrounds and scale variations. Some images may contain artifacts that partially or fully obscure the face.
We conduct extensive experiments using full-body, half-body, and face-only images with different CNN backbones, including VGG, DenseNet, and ResNet. Our results demonstrate that full-body images consistently outperform half-body and facial images for BMI, weight, and height estimation in unconstrained settings.
Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Cite as: arXiv:2607.26104 [cs.CV]
DOI: https://doi.org/10.48550/arXiv.2607.26104
via ArXiv CV