\BeginparThe aim of this research is the direct computation of the sign of Gaussian curvature of pictured objects from images acquired with CCD cameras, without using the classical approach of computing object surface normal vectors. To achieve this, the authors compute the sign of curvature in photometric space with triplets of intensity values for each pixel. The main idea is to exploit the fact that the endpoints of normal vectors along a closed curve \( \SGMPmathgrk{g}\) on a surface \( S\) form another curve \( \SGMPmathgrk{b}\) on the Gaussian sphere, which has only two possible orientations.\Endpar \BeginparThis paper will interest researchers in computer vision who are involved with scene reconstruction from photometry. The method presented avoids a number of recurrent problems with Gaussian curvature computations. For instance, computation of normal surface vectors as an inter\-mediate step is avoided altogether, and the method is albedo-invariant. The accuracy of results is only loosely coupled with the location of light sources, and no calibration of equipment is required. Moreover, no assumptions are made about the type of diffuse reflectance. A possible drawback of the approach is its intrinsic inability to handle objects with diffuse and specular reflectance. However, this problem can be overcome by using cross-polarization, which greatly reduces specular components, as the authors point out. It would have been interesting if the authors had compared their approach with those of others in the field. However, this omission can hardly be seen as a drawback, given the quality of the contribution. Convincing numerical results are presented.\Endpar (Provider: ACM)
Reviewer: Steven S. Beauchemin (Philadelphia, PA)