Grid'5000 user report for
User information( user)
More user information in the user management interface.
- Solving the SFAS problem in the video card (Other) [achieved]
Description: At STEEP, we work on some computer vision problems, being SFAS (Shape From Ambient Shading) one of those. This problem consists in recover the 3D shape of a surface given only one B&W image of the surface illuminated with "ambient light". We already have a theoretical solution for a particular variation of this problem and we are extending it to other situations. We need to run several tests with different surfaces to evaluate the convergence speed and quality of this new approximation. We also have an implementation of this new solution written in C, using off-screen OpenGL rendering and MPI to recover a single surface across several computers without a noticeable performance degradation up to 16 machines.
Results: After several tests using the adonis cluster we found that OpenGL approach to compute the visibility cone in a surface has some problem. First of, there are several technical problems that don't permit the use of the two video cards of the same machine at the same time for the OpenGL offscreen rendering (simply the driver doesn't work after try it with several versions and configurations, lending a non working pbuffer, or a extremely slow rendering time, even worse than using just one video card). After this technical problems, there are also numerical problems given by the fact that we are computing an integral as the sum of so many terms computed by the GPU. And finally, there's also a trade-of between the speed and the accuracy of the result governed by the size of the rendered image. Provided this, the quality of the result and the algorithm's speed weren't good enough to test our solution to this problem. Thus, we change the implementation for a completely different approach to compute the VC, using the CPU in this case (see next experiment).
- Solving the SFAS problem in the CPU (Other) [in progress]
Description: At STEEP, we work on some computer vision problems, being SFAS (Shape From Ambient Shading) one of those. This problem consists in recover the 3D shape of a surface given only one B&W image of the surface illuminated with "ambient light". In this case, we developed a new way to compute the visibility cone and we use it to apply the ideas from other SFS problems to the SFAS problem.
Results: Preliminary results show that this implementations performs better than the OpenGL approach to this problem, in terms of both speed and accuracy, with a better usage of the resources.