It can be seen from the experimental data, that the temporally coherent algorithm can decrease the processing time needed to produce acceptable results. The results suggest that as the amount of knowledge about the robot motion increases, so does the speed of the algorithm. It should be noted that the significant jump in the performance can be observed when the general direction of the camera motion is specified. This is important because the accurate odometry readings may not be available on all robots, but it is likely that the robot knows that it is moving in a particular direction.
The accuracy of the depth map can be as good or better when using the temporally coherent algorithm. We have presented the number of matches of valid disparities between the full and coherent algorithm. The number of matched valid disparities is above 85%. The coherent algorithm does find pixels to be valid even though the full algorithm finds them invalid. The number of pixels found to be valid only by the coherent algorithm are in the range of 10 to 15 % of all pixels, for more constrained motion. The explanation for the this phenomenon is that the results of stereo are temporally extended when the search range is limited. Therefore the algorithm is still able to identify the disparity as valid. If the disparity range was increased the algorithm would find the disparity invalid.
In the case when the motion of the camera is known, the disparity ranges are reduced to less than 10% of the full disparity range. This would lead us to believe that the algorithm should run 1000% faster. The speedup however is only in the range of 400%. A part of the reason for this unexpected performance is the time spent producing the disparity ranges. However, more important is the fact that the algorithm is trying to take advantage of recursion. The coherent algorithm needs to check if the necessary information is available. By doing this it executes many conditional jumps which are inherently expensive on sequential computers.
The disparity maps presented in this paper have on average 60% of all pixels valid. This is quite high considering that no interpolation was done. This was done on purpose in order to force the coherent algorithm to perform computation. Greater speedups are possible if the obtained disparity maps are very sparse. In this case the algorithm may choose not do process disparities that are believed to be invalid in the next iteration. In this case the speedups can go as high as 1000%. The problem with sparse depth maps is that there must be a mechanism for introducing valid points in the regions where invalid disparities are expected. Otherwise the whole image could possibly turn invalid. As our future work we propose two methods as a solution to this problem: statistical verifying of results and/or use of additional knowledge.
The statistical verification means selecting a number of random pixels and processing them over the full range of disparities. The obtained results are then compared with results of the coherent stereo. If the results are different then that part of the image can be processed with a larger disparity range. The statistical verification can be used for introducing valid points to an area of the image that was previously invalid. It could also be used for detecting dynamic objects in the environment.
The second approach to solving the problem of introducing valid points can be solved by using additional knowledge such as the image content. For example, it is well known that correlation-based stereo algorithms perform poorly on texture-less surfaces. Therefore, the checks for texture in parts of the image where it is necessary can help in deciding on whether computation should be done or not.
Additional information can also be useful depending on the task of the robot. For example, a well calibrated robot can determine disparity ranges that correspond to a particular part of the environment. The floor would be ideal to ignore, given that there are disparities that correspond to points below the floor. On the other hand, the robot may be particularly interested in holes in the ground. In that case the disparities should be tuned to find points below the floor level.