VAT-ABAW

Combine 3D-CNN and Transformer for Valence-Arousal estimation on the AffWild2 dataset.

Getting Started

Install requirements

pip install -r requirements.txt

Prepare ABAW dataset

• First, apply for the newest version Aff-wild2 dataset. URL: https://ibug.doc.ic.ac.uk/resources/fg-2020-competition-affective-behavior-analysis/

• Second, download the cropped-aligned images and annotations, and put them on the folder data/ of this project

• Third, run data/generate_dataset.py, then you should get a folder data/dataset which contains Training_Set and Validation_Set images and annotations, also the Test_Set images.

Training

To train models, change paths in mypath.py with your ones, and run train.py We have done several experiments on different models, I will continue to push them into this repo.

Paper

Will be updated soon.

Contact and information

please contact me: [email protected]