Optic Flow

To compare the performance of the local histograms with existing methods used for expression modeling, we calculated the optic-flow of selected regions. The flow estimate is obtained with a multi-scale coarse-to-fine algorithm based on a gradient approach described by [3,11]. Optic-flow is sensitive to the motion of image regions and the direction in which different facial features move, but it is also sensitive to the positions of the feature points.

We computed the dense optic flow of the images extracted from the normalized face image. Then we calculated the eigenvectors of the covariance matrix of the flow images. The top 20 eigenvectors were used to represent the motion observed in the image regions. These captured 85$\%$ of the variance. The eigencoefficients computed for each image region were the new input features. However, large tracking errors leading to improper normalization can cause optic-flow to provide misleading results as shown in Table 1.