Automated Facial Expression Recognition for Product Evaluation

Facial expression play an important role in signaling interest and opinion about a product. In this paper we present a negative affect classifier that detects nose wrinkles (AU9) or brow lowerer (AU4). Our approach uses an automatic mechanism for extracting the frames of interest in videos. We employ an automatic facial feature tracker to detect the area of interest in each frame, we then apply Gabor Jets for feature extraction followed by applying a dynamic Bayesian network classifier. We validate our classifier on two corpora: the first is of posed videos, the second is from a corpus of spontaneous videos of people evaluating two different beverages. We achieved an average accuracy of 97% for posed videos and 96% for spontaneous videos. We also, show that negative facial expressions captured using our machine vision technique are more predictive than selfreports. This was proven by correlating between self reports and reports provided by human coders which is used as our ground truth for our classifier.