Dataset Card for Action2MotionBehavior

Emotion-based dataset containing videos of 30 students from the Faculty of Mathematics and Computer Science, University of Bucharest, filming themselves expressing one of the six following basic emotions: anger, disgust, fear, happiness, sadness, surprise. The features of this dataset were extracted for the task of action-to-motion (A2M).

Additional Repository Details

• "Feature Extraction" directory: the joints and poses extracted from each video of the dataset, following the SMPL and SMPL-X formats, along with labels (0-5 corresponding to anger-surprise). This was done with the help of the HybrIK-X method, using the "HRNet-W48 + RLE" backbone. Moreover, each video was reformatted beforehand with the help of the FFmpeg command "ffmpeg -i input_path -vf scale=new_res -r 20 -c:v libx264 -preset ultrafast -crf 23 -an" where "new_res" was "iw/2:ih/2" if the video was filmed in 4K, or "iw:ih" otherwise
• "Model Training" directory: the model trained from scratch using only the features from the "Feature Extraction" directory, excluding the SMPL-X ones. This model is MDM

Authors

• Created by: Faculty of Mathematics and Computer Science from the University of Bucharest, Ciprian I. Păduraru, Alexandru C. Sasu, Monica A. Gîrbea
• Cleaned by: Alexandru C. Sasu, Monica A. Gîrbea
• Reformatting, feature extraction, and model training realized by: Alexandru C. Sasu