This chapter presents experimental results that demonstrate the feasibility of the method. Each experiment will consider separate aspects of the implementation, including the effects of parameter variation, lighting variation, and environment scale. Results from different environments will also be presented. Finally, we will briefly consider a method for recovering camera orientation even when training is performed at a fixed orientation.
An issue that is of importance is that of how we measure the accuracy of the localisation method. In practice, the goodness of the results will be tied to the sampling density used for training and hence we express the accuracy of experimental results as a percentage of the sample spacing , measured as the average distance between nearest neighbours in the set of poses used for training. We are striving for results that are accurate to a fraction of . A second issue is the difficulty of measuring sufficiently accurate ground truth in some experiments. This has a particularly important impact on the images obtained for training, since it may not always be possible to ensure that the camera is facing in a fixed orientation or that the position of the camera is precise. Indeed, the quality of the results will hinge to some extent on the precision of the training poses. In our discussion of each experiment, we will consider the precision to which we can measure ground truth and compare the results to this measure. Finally, there are some implementation details which should be noted. Unless otherwise stated, all the images used for training and testing are grey scale, at a resolution of 320 by 240 pixels. The window used for measuring edge density is 15 pixels in radius, and the local maxima of edge density are considered only if they differ from the mean edge density by more than one standard deviation. In the tracking phase, better matches to tracked landmarks are sought out over a 20 by 20 pixel neighbourhood of the candidate under consideration.