Newswise — Estimating height and weight from a single face image is valuable for security and health monitoring. However, real-world images often feature non-frontal poses, where significant variations distort facial geometry and hinder precise feature extraction. Furthermore, labeled data for height and weight are extremely scarce compared to facial recognition datasets, causing high risks of overfitting in high-dimensional regression models. This combination of “pose interference” and “data scarcity” has long remained a primary bottleneck for practical biometric deployment.
To address this, the ICT-CAS team proposed a framework utilizing multi-task learning and pose-disentanglement. By incorporating gender and age estimation as auxiliary tasks, the model leverages biological correlations and more abundant data to enhance feature representation. Crucially, a pose-disentanglement mechanism strips pose-relevant redundant information from input features. This allows the model to focus on robust intrinsic traits related to body metrics, significantly reducing prediction bias caused by head rotation and ensuring stable performance across different perspectives.
Experiments on datasets with diverse poses show that the method significantly outperforms current benchmarks, exhibiting exceptional robustness under extreme variations. Ablation studies confirm that auxiliary tasks effectively mitigate data scarcity, while pose-disentanglement improves generalization across various scenarios. This research provides a technical foundation for multi-modal body analysis and intelligent health monitoring.
https://link.springer.com/article/10.1007/s11704-025-50162-0
https%3A%2F%2Fwww.newswise.com%2Farticles%2Fface-based-height-and-weight-estimation-via-pose-disentanglement%2F%3Fsc%3Drsla
