MIMAMO Net: Integrating Micro- and Macro-motion for Video Emotion Recognition
Paper Link: https://arxiv.org/pdf/1911.09784.pdf
Requirements:
- Pytorch 0.4.1 (or higher version)
- Numpy
- PyTorchSteerablePyramid
- pytorch-benchmarks
- OpenFace
In this paper, we propose to combine the micro- and macro-motion features to improve video emotion recognition, using a two-stream recurrrent network named MIMAMO (Micro-Macro-Motion) Net. This model structure is shown in the picture:
To run this project,
(1) Download the pretrained ResNet50 model from this webpage, which is pretrained on VGGFACE2 and FER_plus. Make sure the pytorch-benchmarks is correctly installed and the pretrained model can be imported.
(2) Use OpenFace toolkit to crop and align faces in videos, save aligned faces.
(3) Extracted the Pool5 features of ResNet50 model and save features. Using the python script in './scripts/CNN_feature_extraction.py':
python CNN_feature_extraction.py --fps 30 --layer_name pool5_7x7_s1 --save_root Extracted_Features --data_root dir-to-aligned-face
(4) Before running experiments on Aff-wild dataset (or OMG emotion dataset), make sure dataset is downloaded and processed in step (3).
Run scripts in 'Aff-wild-exps' or 'OMG-exps'.