New Face-Based AI Algorithm Accurately Estimates Body Height and Weight Using Novel Pose-Disentanglement Technology

New research led by Shiguang Shan uses pose-disentanglement and auxiliary tasks to accurately estimate body height and weight from facial images.

By: AXL Media

Published: Apr 25, 2026, 8:05 AM EDT

Source: Information for this report was sourced from EurekAlert!

Overcoming Visual Distortions in Biometric AI Modeling

Estimating physical body metrics such as height and weight from a single facial image has historically been hindered by the significant variability in head poses and a lack of annotated datasets. To address these limitations, a research group directed by Shiguang Shan has introduced a method that isolates facial features from the physical orientation of the head. This process, known as pose-disentanglement, allows the AI to focus on stable biological indicators rather than the angle at which a photo was taken. By removing pose-relevant data from the primary feature set, the algorithm can more accurately identify the subtle facial markers associated with an individual's overall body mass and stature.

The Correlation Between Age Gender and Physical Stature

The strategic breakthrough of this research lies in the use of auxiliary tasks, specifically the estimation of age and gender, to refine the primary predictions of height and weight. The researchers noted that body shape and bone density evolve predictably over time, meaning age-related features provide essential context for physical dimensions. Furthermore, because males and females typically exhibit different ratios of muscle mass and bone mineral density, gender perception serves as a critical weight-filtering mechanism. According to the study, integrating these related biological attributes into a single learning model significantly boosts the performance of the height and weight estimations compared to models that view these metrics in isolation.

Architectural Innovations in Multi-Task Feature Learning

The technical framework begins with a series of convolutional layers designed to extract general facial features from an input image. These features are then passed through a disentanglement module that filters out data related to head rotation and tilt, leaving behind a pose-irrelevant feature set. Subsequent branches of the network learn task-specific details for age and gender, which are then fused back into the main pipeline. This fusion process ensures that the final convolutional and fully-connected layers have a comprehensive profile of the subject. The entire system is optimized using multi-task losses, allowing the AI to learn from the interdependencies between different human attributes simultaneously.