We have demonstrated the ability of our method to accurately measure classes of motion, such as translation in depth, rotation out of the image plane, and large angle rotation, which have caused difficulties for most previously described techniques. We have successfully applied the technique to real imagery, and have shown that in practice, the linear brightness and depth constraints complement each other. Finally, we were able to track motion with very little cumulative error over long video sequences (up to 10 seconds, or 300 frames).
One conclusion that we can draw from the above experiments is that the use of the combined intensity and depth constraints outperforms the use of either independently. This occurs not only because the two constraints together provide more data upon which to base estimates, but also because each type of constraint helps offset the shortcomings of the other. For example, the DCCE is relatively insensitive to photometric effects, while the BCCE typically deals with data that is less noisy and has fewer undefined regions. Another conclusion we can reach is that the substitution of measured, frame-rate depth for generic object shape models can substantially improve on pose estimation methods which use only the BCCE or variations on it. The same is likely true for methods which estimate object shape and motion together.
Our method lends itself nicely to real-time systems, in that it is a linear method which may be solved efficiently via least-squares. Another advantage of our method is that, because it is a differential tracker that updates its shape model of the object over time, it can track object pose through dramatic changes such as those shown in the synthetic head rotation sequence. In fact, we should expect the method to perform reasonably well in tracking an object of arbitrary shape through full 360 degree rotations, assuming small inter-frame motions.