I was a bit surprised at the lack of detail concerning the implementation, particularly the results, in the Lucas paper. After all the time spent deriving the system and the indepth explaination, I would have liked to get a slightly better feeling for how well the system worked.

Along these same lines, I also would have liked to have some more concrete examples of what applications this is commonly used.

I was also a bit surprised at how quickly the paper dismissed the idea of using segmentation to help results. However, upon further consideration, given the time that this paper is published, I think segmentation was no where near the performance that it is now (which is still by no means perfect).

An obvious problem with the flow field algorithm is the smoothness assumption. The world with multiple objects moving will certainly conflict with that assumption, and there are a lot of effort made in relexing the constraints. However, a funny thing is that the brain area that detect 2D movement (area MT) has very large receptive fields. Since MT neurons are integrating movements in a large area, MT representation is very smooth - much like the Horn paper. The puzzle is how can such a smooth representation be helpful at all to the visual system? Last time I checked there is no answer for that yet.

I want the detail information of the 3D recovery by using the optiacl flow.please help me .This is something which is very important for me.

Im looking for informations about optical flow.
Especialy for the navigation of indoor roboters.
It would be very great if can retrieve some links (for instance of publications).

Thank you all..