An accurate, real-time algorithm for the automatic annotation of a million facial expressions in the wild
A new computer vision al- gorithm for the recognition of AUs typically seen in most applications, their intensities, and a large number (23) of basic and compound emotion categories across databases. Additionally, images are annotated semantically with 421 emotion keywords.
Link to dataset : http://www.pitt.edu/~emotion/downloads.html