Processed for a Neural Network

AffectNet is a large database of faces labeled by "affects" (psychological term for facial expressions). In order to accommodate common memory constraints, the resolution was reduced down to 96x96. Meaning that all images are exactly 96x96 pixels.

No Monochromatic Images

Using Singular Value Decomposition, each image's Principal Component Analysis was calculated. The threshold for the "percentage of the first component (index 0) in the principal components" (in short the PFC%) was set to lower than 90%. This means that most if not all of the monochromatic images were filtered out. The PFC% was left inside the CSV file, should you prefer a lower PFC% threshold.

Pseudo labels

This data-set is based on AffectNet-HQ, which used a state-of-the-art FER model to improve the on the AffectNet original labels.