The stereo algorithm starts by computing the Laplacian of the Gaussian for each of the three images. The algorithm then loops over the disparities and computes the sum of the absolute values of the differences (SAD) between the image pairs using a 5 by 5 window. It keeps track of the disparity with the minimum SAD for each pixel. The correlation and filtering is done using two Thompson A110 convolver chips that run in parallel with the C40.
The depth image computed from the scene in Figure 3 is shown in Figure 5. Points that are closer in the image are displayed as a lighter color of gray and invalid points are displayed as completely black. This stereo data is dense and has almost no noise.
