The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
792 
seven images were collected by using the same camera in 
different locations. Although only one camera was used in our 
experiments at the current stage, seven cameras should be used 
in the future to ensure simultaneous data acquisition. The three 
following datasets were acquired: 
1.1 mages of the first person at time T ]. (David_ 1) 
2. Images of the second person at time T 2 . (Ivanl) 
3. Images of the first person at time T 3 . (David_2) 
The pattern projection produced encoded patterns on the face 
and provided specific landmarks on homogenous surfaces, 
thereby solving the problem of identifying conjugate features in 
the stereo-pair. To increase the reliability of the matching 
process, the internal and external characteristics of the utilized 
cameras and images, respectively, were incorporated to reduce 
the search space from 2D to ID by using epipolar geometry for 
stereo-imagery. After the Harris operator was applied to 
automatically detect comers resulting from the pattern 
projection system, the NCC was then performed for identifying 
conjugate features between adjacent images (Figure 4). 
Figure 4. Extracted comers and detected tie points on images 
#1 and #2 in Ivan_l. 
Following the matching process, the tie points were tracked 
through all images captured in different locations. Using 
tracked conjugate points between two images with an 
intersection angle close to 90 degrees for best accuracy, an 
intersection procedure was performed to derive the 3D 
coordinates of the corresponding object points. A TPS 
algorithm was then used to generate a 3D facial model with a 
dense set of points. Figure 5 shows the reconstruction results 
from conjugate light rays intersected by the stereo images with 
the largest baseline after tracking, while the surface shown in 
Figure 6 was reconstructed only using one pair of adjacent 
images. Compared with the noisy mesh produced without a 
tracking process in Figure 6, the accuracy of the output meshes 
in Figure 5 is much superior. 
Figure 6. Surfaces reconstructed using stereo-pair with a short 
baseline. , 
Figure 7. Co-registered facial models using Davidl and 
David_2 with 93.902% matched points. The green mesh 
represents the facial surface of Davidl, and the points 
represent the facial surface of David_2 (blue: matches, red: 
non-matches). 
Figure 8. Co-registered facial models using David l and 
Ivan l with 79.458% of matched points. The green mesh 
represents the surface from David l, and the points represent 
the surface from Ivan l (blue: matches, red: non-matches). 
Figure 5. Surfaces reconstructed using stereo-pair with a wide 
baseline.