CUDA-based Reconstruction and Visualization of Non-uniform Point Sets

Information

• Publication Type: Student Project
• Workgroup(s)/Project(s):
  • Visualization Group
• Date: 2011
• Matrikelnummer: 625909
• First Supervisor: Erald Vucini
• Keywords: CUDA, GPU, Variational Reconstruction

Abstract

In this work we present a CUDA-based parallelized implementation of the reconstruction technique for 2D and 3D non-uniform point sets using a variational approach. Cubic B-splines are used as an approximation for the computationally expensive radial basis functions. As B-splines only have local support, a regularization term is added to handle larger gaps between sample points. The main reconstruction task is the solving of a large but sparse linear system. The system's size makes the use of iterative algorithms necessary. Different solvers are implemented in order to find the best-performing approach. GPU-based algorithms require a different kind of strategy than traditional implementations. The possible degree of parallelization becomes the defining factor for the efficient realization of algorithms. The use of B-splines also makes a multi-resolution approach possible. The system can be solved at a coarser representation and its solution acts as an initialization for the finer resolution. This further speeds up the reconstruction. A downside of using graphics cards for computation is the relative sparsity of available memory. Most GPUs are not capable of handling big datasets directly. Therefore, the non-uniform representation is split into a number of smaller blocks. Each block is then reconstructed separately and the intermediate results are joined together resulting in the final solution. The reconstruction results in terms of speed and quality are shown and are pitted against other CPU-based implementations. A CUDA based interactive ray-tracer is used to visualize the 3 dimensional datasets and test the visual quality of the reconstruction.

Additional Files and Images

Weblinks

No further information available.

BibTeX

@studentproject{Csuk_2011,
  title =      "CUDA-based Reconstruction and Visualization of Non-uniform
               Point Sets",
  author =     "Maximilian Csuk",
  year =       "2011",
  abstract =   "In this work we present a CUDA-based parallelized
               implementation of the reconstruction technique for 2D and 3D
               non-uniform point sets using a variational approach. Cubic
               B-splines are used as an approximation for the
               computationally expensive radial basis functions. As
               B-splines only have local support, a regularization term is
               added to handle larger gaps between sample points. The main
               reconstruction task is the solving of a large but sparse
               linear system. The system's size makes the use of iterative
               algorithms necessary. Different solvers are implemented in
               order to find the best-performing approach. GPU-based
               algorithms require a different kind of strategy than
               traditional implementations. The possible degree of
               parallelization becomes the defining factor for the
               efficient realization of algorithms. The use of B-splines
               also makes a multi-resolution approach possible. The system
               can be solved at a coarser representation and its solution
               acts as an initialization for the finer resolution. This
               further speeds up the reconstruction. A downside of using
               graphics cards for computation is the relative sparsity of
               available memory. Most GPUs are not capable of handling big
               datasets directly. Therefore, the non-uniform representation
               is split into a number of smaller blocks. Each block is then
               reconstructed separately and the intermediate results are
               joined together resulting in the final solution. The
               reconstruction results in terms of speed and quality are
               shown and are pitted against other CPU-based
               implementations. A CUDA based interactive ray-tracer is used
               to visualize the 3 dimensional datasets and test the visual
               quality of the reconstruction.",
  keywords =   "CUDA, GPU, Variational Reconstruction",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2011/Csuk_2011/",
}