Assessing the performance of the full algorithm is straight forward. The speed at which the full algorithm executes depends on the resolution of the image, the size of the correlation mask and, most importantly, the size of the fixed disparity range. The execution time of the full algorithm is linear to the number of pixels and disparity range and constant with respect to the size of the correlation mask (due to recursion).
The performance of the temporally coherent stereo algorithm on the other hand is quite complex. The complexity arises from the fact that the performance depends on the additional information available to the algorithm.
The first factor is the knowledge of motion. If the motion is known, then it is possible to accurately estimate the disparity search ranges. If the disparity ranges are small then the stereo search will be done faster.
The second factor is the structure of the scene. If the scene consists mostly of flat surfaces, then the predicted disparity ranges will be around the previous values. If, on the other hand, the scene has many discontinuities, then the search at the discontinuities will have to include both ranges which may be far apart.
The third factor is the number of valid and invalid pixels in the image. The algorithm processes only parts of the image that may move to a valid part of the image. If the image has large areas of invalid pixels and the motion is well known or small, then it is possible to determine areas of the image that do not need to be processed. This means that images that lead to large numbers of invalid disparities will execute faster with the coherent stereo algorithm.
Another important point is that even if the disparity ranges can be determined precisely, the program can still perform inefficiently. The inefficiency is due to the recursive nature of the algorithm. In other words, if one correlation can not benefit from the previously done work, the computation is done inefficiently. Scattered invalid points and small areas of different disparities can cause this effect.
Finally, close objects cause greater change in the disparity values when the robot moves towards or away from them. Therefore the algorithm will have poorer performance when objects are in close proximity.
The performance of the stereo algorithm, when given the disparity ranges, is one part of the computational cost of the whole algorithm. The other part is determining the disparity ranges. The time required to compute the disparity ranges depends mainly on the amount of robot motion. In general the less the knowledge of motion is constrained the longer it takes to compute the disparity ranges. Loose constraints on motion mean that larger ambiguity areas, which result in more searching for the disparity range.