DiffME: A Diffusion-based Unified Framework with Soft Pseudo-Labeling for Micro-Expression Spotting and Recognition

Official PyTorch implementation of "DiffME: A Diffusion-based Unified Framework with Soft Pseudo-Labeling for Micro-Expression Spotting and Recognition" (CVPR 2026 Workshop).

---

📌 Overview

DiffME is a novel diffusion-based unified framework designed for joint Micro-Expression (ME) spotting and recognition. By incorporating soft pseudo-labeling, DiffME effectively addresses the challenge of subtle facial movements and imbalanced, noisy annotations in micro-expression datasets such as SAMM Long Videos and CAS(ME)³.

---

🔧 Setup

STEP1: bash setup.sh

STEP2: conda activate DiffME

STEP3: pip install -r ./requirements.txt

💻 Examples of Neural Network Training

STEP 1: Download the SAMMLV/$CAS(ME)^3$ raw data by asking the paper authors

STEP 2: Modify main.py; load_excel.py; load_images.py

STEP 3: Run python main.py --dataset_name CASME_3 --train True

🎓 Acknowledgement

We referred to ME-TST+, and would love to thank them for having shared the code publicly.

📜 Citation

If you find this repository helpful, please consider citing:

@InProceedings{Ha_2026_CVPR,
    author    = {Ha, Pham Ngoc Thach and Nguyen, Luu Tu and Ha, Le Thanh and Ngo, Thi Duyen},
    title     = {DiffME: A Diffusion-based Unified Framework with Soft Pseudo-Labeling for Micro-Expression Spotting and Recognition},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2026},
    pages     = {1310-1319}
}