A series of corner detections is performed on each image taken, with the window size used to calculate the exact corner points varying. These results are then extrapolated down to an infinitesimal window size; in theory, this will produce the corner positions as exactly as they can be computed. In practice, a window size of one or three pixels squared seems to be better than zero for estimating positions. Since the coordinate adjuster does not require a certain window size, this doesn't matter too much, but it would be nice for the theory to fit experimental data-i.e., for the results given by the data from a window size of 0 to fit better than the results given by a window size of 1 pixel.
Different window sizes are used to calculate sets of corner points because the larger the window used, the more gradient data is given, but the smaller the window used, the more exact the analyzer can be. In addition, this type of routine should be able to be used on very small images that cannot support large window sizes. However, the smaller the window used, the less data is available, and at a window isze of about 1 pixel, there is no variation in image data, the gradient is useless and no routine can exactly find any corner based on zero data. This project does not attempt to answer the question of what window size is most useful in detecting corners.