During the january 28 lecture, we intend to cover:

E. Chen, L. Williams, View Interpolation for Image Synthesis, SIGGRAPH 1993
S. Seitz, C. Dyer, View Morphing SIGGRAPH, pp. 21--30, 1996
Y. Genc, J. Ponce. Image-Based Rendering Using Parameterized Image Varieties. International Journal of Computer Vision, Vol. 41, No. 3, pp. 143-170, 2001.

It seems that the view interpolation paper was motivated somewhat by the lack of effective hardware for real-time rendering. They say that the scene complexity problem "is potentially unbounded" and will exist independent of the processing power of the computer. I don't know if the authors could have foreseen the advancement in graphics hardware that allows for very complex scenes to be rendered.

Here's a question:
Do you think that scene complexity is still such a big concern, and the main problem that needs to be dealt with?

Hi Peter,

The scene complexity might not be a big issue in itself with the advancement of graphics hardware. But, what the author wants to say is that, an algorithm shouldnt ideally depend on scene complexity which can be arbitrarily complex. So, no matter how fast graphics hardware is used, there is always a finite probability that the scene wont be rendered as fast as we want. One of the key words in definition of realtime systems is " Predictability". Hence,
it is essential that the timing constraints of the system are guaranteed to be met. If a rendering algorithm is scene dependent, no guarantee can be made about real time performace. For the sake of completeness, here's a classical canonical definition of real time systems:

"A real-time system is one in which the correctness of the computations not only depends upon the logical correctness of the computation but also upon the time at which the result is produced. If the timing constraints of the system are not met, system failure is said to have occurred."

Others have added:

"Hence, it is essential that the timing constraints of the system are guaranteed to be met. Guaranteeing timing behavior requires that the system be predictable. It is also desirable that the system attain a high degree of utilization while satisfying the timing constraints of the system."

The view interpolation paper posted on the class website is incomplete. It has references to figure which donot exist. Please go to the acm digital libary website to get a "complete" copy of the paper. Caution: The size of the document is huge(2 MB).

The Seitz and Dyer paper assumes the availability of dense correspondence besides the availabiity of the projection matrices and yet they claim they do not require structure. If you have those two things correctly, you infact have the structure of the scene. It is a simple calculation away. Hence the whole thing about how they are able to do this interpolation without recovering structure is quite bogus.

Hi Sameer,

Although, they say in the paper about dense stereo, but I am quite sure, they are not actaully doing dense stereo. They are probably doing standard trinagulation based morphing after rectification. I think their main contribution was to make the morphing "Physically Valid"

I guess the reason I brought up the matter about scene complexity is because it seemed like the authors were focusing on scenes where unnaturally good information is available (3D pixel values). The application is thus suited more toward handling in-betweens and novel viewpoints of a computerized model. Nowadays, we'd do that with our video hardware.