A comparison of detailed automatic wire-frame fitting methods

This paper presents a comparative study of accurate automatic semantic wire-frame fitting methods for extremely low bit-rate model-based moving image communication systems. Efforts are concentrated on the accurate fitting of the semantic wire-frame in areas occupied by important facial features (the lips, the eyes and the nose). The automatic fit in the areas occupied by those features is crucial for the correct reconstruction of the transmitted scene. The algorithms are based on the analysis of a facial code-book and are designed to provide the final and most accurate fit of the Candide wire-frame model to the facial features of the speaker in a typical videophone scene (head-and-shoulders). The results obtained from tests carried out on the images from the MIT facial data-base are presented and compared