PhysDiff: Physiology-based Dynamicity Disentangled Diffusion Model for Remote Physiological Measurement
Authors: Wei Qian, Gaoji Su, Dan Guo, Jinxing Zhou, Xiaobai Li, Bin Hu, Shengeng Tang, Meng Wang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on four datasets demonstrate that our Phys Diff significantly outperforms prior methods on both intra-dataset and cross-dataset testing. |
| Researcher Affiliation | Academia | 1School of Computer Science and Information Engineering, Hefei University of Technology 2Institute of Artificial Intelligence, Hefei Comprehensive National Science Center 3School of Cyber Science and Technology, Zhejiang University 4School of Information Science and Engineering, Lanzhou University EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the model architecture and processes (Forward Process, Reverse Process, Dynamicity Disentanglement, Spatial-Temporal Hybrid Denoiser) using mathematical formulations and textual explanations. However, it does not include any explicitly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | Code https://github.com/VUT-HFUT/Phys Diff |
| Open Datasets | Yes | PURE (Stricker, M uller, and Gross 2014) recorded a total of 60 videos featuring 10 subjects across six different scenarios. UBFC-r PPG (Bobbia et al. 2019) contains 42 videos recorded in a stable laboratory scenario. VIPL-HR (Niu et al. 2019) is a challenging dataset for remote physiological measurement, which records 2,378 facial videos from 107 subjects under 9 complicated and di-verse scenarios. MMSE-HR (Tulyakov et al. 2016) consists of 102 videos recorded by 40 subjects from 40 subjects of different races with diverse facial expressions. |
| Dataset Splits | Yes | To present a thorough analysis of the proposed method, we conduct detailed ablation studies on fold-1 of the VIPL-HR dataset as the protocol in (Yu et al. 2023; Qian et al. 2024b). |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory amounts used for running the experiments. It only mentions the implementation framework. |
| Software Dependencies | No | The proposed Phys Diff is implemented in Py Torch using the Adam optimizer. |
| Experiment Setup | Yes | The learning rate is set to 1e-3. We train our model for 50 epochs on each dataset. During the training stage, the number of hypotheses and iterations H, K is set to 1,1, respectively. During the inference stage, they are set to 10 and 5. The maximum diffusion steps T is set to 1000. The loss weight λ in Eq. 10 is set to 0.1. The depth of denoiser is set to 6. |