as 
e 
id 
2) 
re 
e, 
on 
Fua, Pascal 
  
  
  
  
(d): 
Figure 10: Reconstruction and animation. (a) For the subject of Figure 8(a): The manually supplied 2-D features points; two shaded 
views of the complete head model; the cylindrical texture map; and, textured model. (b) For the subject of Figure 8(b), the 
manually supplied 2-D features points; two shaded and two textured views of the head model. (c) Man with a beard: two 
texture-mapped views of the corresponding head model; and two synthetic expressions, opening of the mouth and closing 
of the eyes. (d) Man with a mustache: One shaded and three texture-mapped views of the corresponding head. 
Shape and Texture Figure 10 depicts the output of our modeling procedure. In all cases, the head models can be 
animated. To ensure that some of the key elements of the face—corners of the eyes, mouth and hairline—project at 
the right places, we have manually supplied the location of the projection in one image of a few feature points such as 
the ones shown in the first columri of Figure 10: We add to the objective function of Equation 2 a term that forces the 
projection of the generic mask's corresponding vertices to be close to them (Fua and Miccio, 1998). Note that the five 
manually supplied points used to initialize the bundle-adjustment procedure of Section 4.1.1 form a subset of these feature 
points. To produce these face models, the manual intervention required therefore reduces to supplying these few points 
by clicking on their approximate locations in one single image, which can be done quickly. 
Because stereo tends to fail in the hair regions, the shape of the top of the head has been recovered by semi-automatically 
delineating in each image of the video sequence the boundary between the hair and the background and treating it as a 
silhouette that constrains the shape (Fua and Miccio, 1998). Given the final head model, the algorithm creates a cylindrical 
texture map, such as the one shown in the top row of Figure 10. 
For a quantitative evaluation of these results, we have acquired 3-D models of the faces of both subjects of Figure 8 
using a Minolta” scanner. The theoretical precision of the laser is approximately 0.3 millimeters and it can therefore be 
considered as a reasonable approximation of ground truth. In practice, even the laser exhibits a few artifacts. However, 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 263