We developed EmoDetect, an open source system for identifying human emotion from images. This system has an accuracy of around 63% in identifying the correct emotion from amongst seven candidates. Our experiments showed that humans did slightly better, with an accuracy of 74%.
The project was aimed as a comparative study of different feature extractors and learning algorithm combinations for the task of identifying emotion from static images. Our study revealed that Gabor features with a linear, soft margin SVM performed best.