Abstract
To build a robust system for predicting emotions from usergenerated videos is a challenging problem due to the diverse contents and the high level abstraction of human emotions. Evidenced by the recent success of deep learning (e.g. Convolutional Neural Networks, CNN) in several visual competitions, CNN is expected to be a possible solution to conquer certain challenges in human cognitive processing, such as emotion prediction. The emotion wheel (a widely used emotion categorization in psychology) may provide a guidance on building basic cognitive structure for CNN feature learning. In this work, we try to predict emotions from user-generated videos with the aid of emotion wheel guided CNN feature extractors. Experimental results show that the emotion wheel guided and CNN learned features improved the average emotion prediction accuracy rate to 54.2%, which is better than that of the related state-of-the-art approaches.