Stereo algorithms can provide useful three-dimensional information for
a variety of purposes, including face recognition. For good accuracy,
the distance between cameras should be large. However, the large base
line often makes the correspondence problem more difficult. In this pa
per, we examine the problem of non-Lambertian reflectivity in multi-ba
seline stereo and propose a refined multiple-baseline stereo technique
which estimates the brightness on each camera by using adjacent camer
as. In this way, the erroneous stereo correspondences due to brightnes
s changes caused by non-Lambertian reflectance can be reduced without
needing to explicitly model the actual reflectance model. Experimental
results are presented and significant reductions in correspondences e
rrors are demonstrated in comparison with previous methods. (C) 1997 E
lsevier Science B.V.