AI Model Estimates Weight and Height from a Single Photo

A team of computer scientists from Pakistan has developed a deep learning model capable of estimating a person's weight, height, and Body Mass Index (BMI) from a single photograph taken in uncontrolled environments. The work, led by Hira Yaseen, Arif Mahmood, and Waqas Sultani, addresses the challenge of automatic BMI estimation from 'in the wild' images—photos with varied poses, backgrounds, camera angles, and partial occlusions common on social media. To train and evaluate their model, the researchers compiled a new dataset of 6,105 full-body and half-body images with verified height and weight labels, covering diverse ethnicities, ages, and genders. They experimented with several deep neural network architectures (VGG, DenseNet, ResNet) and input modalities including RGB, depth maps, pose-affinity maps, and edge maps, employing both single-task and multi-task learning.

The study found that full-body images significantly outperformed half-body or face-only inputs, suggesting that body shape and proportion cues are critical for accurate physical estimation. Among the modalities tested, combining RGB with depth and pose maps improved predictions, though the paper notes that depth maps estimated from single images introduce noise. The best results came from multi-task learning using RGB plus estimated depth on full-body images. This work has potential applications in health monitoring, fitness tracking, and even forensic analysis, where estimating physical characteristics from available photographs could provide valuable insights without requiring in-person measurements.