Similarity – pixels should look alike (image patch on the left should look like the image patch on the right).
Assumptions:
Most scene points visible in both views.
Image regions for matches are similar in appearance.
Tactics- Do a dense correspondence by first:
Normalize images to same standard deviation of pixels, so there are no intensity issues.
Scan through the images comparing every pixel or every window around the pixel in the left image with a pixel or a window around the pixel in the right image by using Compute a matching cost using SSD (sum squared differences) and find the best score/match of least error.
More correspondence soft constraints:
Uniqueness – no more than one match of the left pixel in the right image
Ordering – pixels go the same direction in both images
Disparity gradient is limited – depth doesn’t change too quickly.
