MEB aims to simplify, standardize and streamline the process of micro-expression analysis. It provides tools for data loading and training micro-expression models.
[ArXiv] Data Leakage and Evaluation Issues in Micro-Expression Analysis
The following example shows how to run a pretrained ResNet18 on the cross dataset protocol using optical flow as input.
from meb import core, datasets
from functools import partial
from timm import models
# Load in the data
cross_dataset = datasets.CrossDataset(resize=112, optical_flow=True)
# Define configurations. Inherit core.Config for default configs
class ResNetConfig(core.Config):
model = partial(models.resnet18, num_classes=len(core.Config.action_units), pretrained=True)
# Create a validator for cross-dataset protocol based on config
validator = core.CrossDatasetValidator(ResNetConfig)
# Train and test with the cross-dataset protocol
validator.validate(cross_dataset.data_frame, cross_dataset.data)Experiments for jupyter notebook demos of using the library.
See Getting started for installing and adding datasets. See Config, Datasets and Validation for understanding the pipeline.
git clone https://github.com/tvaranka/meb
cd meb
pip install -e .@article{meb,
author = {Varanka, Tuomas and Li, Yante and Peng, Wei and Zhao, Guoying},
title = {Data Leakage and Evaluation Issues in Micro-Expression Analysis},
journal={arXiv preprint arXiv:2211.11425},
year = {2022},
}