First page Back Continue Last page Overview Text
Local approach to stereo correspondence is one of the oldest and still most common for all real stereo correspondence systems that I’ve seen. To find the disparity for pixel p in the left image, we take a patch or window of points around it. We shift this window to some location, say location corresponding To disparity 1 in the right window. We compute the cost function between these 2 windows. A common cost function is as follows. If these are 2 windows, one from the left and the other from the right images, then the cost between these 2 windows is just the sum of squared differences between the colors of corresponding window pixels. This cost favors windows with corresponding pixels having the same color. This cost is commonly called SSD. Then we repeat shifting the window around pixel p to several other locations, and finally we choose disparity for pixel p, denoted by d_p, which gives the best cost. Most method use a window of fixed size and shape, that is the same window is used for all pixels in the left image.