Weight and Height Estimation from a Single Human Image Captured in the Wild

This paper addresses the challenge of automatically estimating a person's weight, height, and Body Mass Index (BMI) from unconstrained, "in the wild" images. Unlike previous studies that relied on controlled environments or limited facial data, this research aims to build a robust system capable of handling the high variability found in social media imagery, such as diverse poses, camera angles, occlusions, and varying backgrounds.

The researchers curated a new, large-scale dataset, the "Body2BMI-ITU" dataset, containing 6,105 images collected from platforms like Reddit and Imgur. These images include a wide range of subjects, poses (e.g., sitting, side-profiles, selfies), and environmental conditions. The authors employed deep neural networks—specifically VGG-16, DenseNet-121, and ResNet-50—to perform both single-task and multi-task learning. They experimented with various input modalities, including RGB images, depth maps, pose-affinity maps, and human segmentation masks, to determine which features most effectively capture body composition.

The study found that full-body images consistently outperform half-body or facial images for BMI and weight estimation, confirming that the inclusion of body shape and limb proportions is critical for accuracy. Furthermore, the authors demonstrated that multi-task learning—where the model simultaneously predicts height, weight, and BMI—yields superior performance compared to training separate models for each attribute. This suggests that the shared features between these physical characteristics provide complementary information that improves the overall predictive capability of the network.

Accurate, non-invasive BMI estimation has significant potential for public health monitoring, allowing for large-scale analysis of health trends without the need for clinical equipment. Beyond health, the ability to estimate physical attributes from images has practical applications in forensics, security, and person re-identification. By releasing a diverse, publicly available dataset, this work provides a foundation for future research into computer vision applications for human physical analysis.