PaGe: A Physics-Aware Generative Network for Pressure Map Synthesis

作者：Yunkang Zhang, Quan Wan, Changhai Ma, Z W Wu, Chenfang Fang, Xiaohui Cai · 年份：2026 · DOI：10.1109/icassp55912.2026.11462463 · 研究领域：Machine Learning in Materials Science、Advanced Sensor and Energy Harvesting Materials、Neural Networks and Reservoir Computing

Pressure maps, which capture the distribution of force on a contact surface, play a critical role in clinical monitoring, sports science, and human–computer interaction. While pressure information is robust to visual obstacles and lighting conditions, existing acquisition methods often rely on expensive hardware and suffer from limited scalability. Since human meshes can be efficiently obtained and inherently encode posture and contact geometry, they offer a promising surrogate for estimating pressure without physical sensors. To this end, we propose PaGe, a physics-aware generative framework that synthesizes high-fidelity pressure maps directly from human meshes. Built upon a conditional variational autoencoder (cVAE), PaGe introduces a Condition-Modulated Upsampling Module (CMUM) that adaptively modulates features through style-inspired transformations, while leveraging an auxiliary point cloud reconstruction task to further enrich feature learning. Extensive experiments on three public datasets demonstrate that PaGe surpasses state-of-the-art methods in both quantitative accuracy and physical plausibility. Moreover, PaGe facilitates the creation of large-scale synthetic pressure datasets, providing a scalable alternative to hardware-based solutions.