Real-time active shape models for face segmentation

Abstract

In this work we tackle the problem of real-time alignment of active shape models to new object instances at video frame rate. To achieve this we use edge orientation information as the basic image feature. Unlike in the original active shape framework we incorporate the image features directly into the model vector. We also introduce a new update rule for a model point in a local surrounding. We demonstrate the effectiveness of the proposed approach with a face segmentation task. There we are able to fit a new model face on average within 20 msec on 500 MHz Pentium II PC if the initial model position and size does not deviate too much from the true position. The latter is assured by a face detection step which is carried out before the active shape alignment.