.png){kind=icon}
.png){kind=icon}
Return
Phys-Vim: State space model for remote physiological measurement
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.131405.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Remote photoplethysmography (rPPG) is a contact-free and cost-effective method for video physiological signal estimation. Traditional networks in rPPG like CNNs are limited by local inductive bias while transformers require large datasets for training and have quadratic computational complexity. Moreover, existing methods often struggle to accurately extract periodic signals in long- range contexts, especially in the presence of noise or unseen scenarios. To address these challenges, we introduce Phys-Vim, a novel method based on the Vision Mamba (Vim), an advanced state space model for vision tasks. Phys-Vim can capture long-range dependencies with linear computational complexity. Additionally, we propose a Multi-Path Selective Scan (MPSS) mechanism to improve the consistency and global awareness for signal extraction from video. Extensive evaluations demonstrate that Phys-Vim achieves state-of-the-art performance. In the cross-dataset evaluation from UBFC-rPPG to PURE, Phys-Vim reduces MAE by 73 % and increases Pearson correlation by 8 % compared to previous best results. Moreover, it requires far fewer parameters and reduces computational cost by 11.5 % compared to the previous transformer-based Spiking-PhysFormer.
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W