Information

  • Publication Type: PhD-Thesis
  • Workgroup(s)/Project(s):
  • Date: April 2021
  • Duration: 5
  • 1st Reviewer: Mario Botsch
  • 2nd Reviewer: Carsten Dachsbacher
  • Rigorosum: 7. April 2021
  • First Supervisor: Michael Wimmer
  • Keywords: point cloud rendering, lidar

Abstract

Laser scanning, photogrammetry and other 3D scanning approaches generate data sets comprising millions to trillions of points. Modern GPUs can easily render a few million and up to tens of millions of points in real time, but data sets with hundreds of millions of points and more require acceleration structures to be rendered in real time. In this thesis, we present three contributions to the state of the art with the goal of improving the performance as well as the quality of real-time rendered point clouds.

Two of our contributions address the performance of LOD structure generation. State-of-the-art approaches achieve a throughput of up to around 1 million points per second, which requires users to wait minutes even for smaller data sets with a few hundred million points. Our proposed solutions are: A bottom-up LOD generation approach that creates LOD structures up to an order of magnitude faster than previous work, and a progressive rendering approach that is capable of rendering any point cloud that fits in GPU memory in real time, without the need to generate LOD structures at all. The former achieves a throughput of up to 10 million points per second, and the latter is capable of loading point clouds at rates of up to 37 million points per second from an industry-standard point-cloud format (LAS), and up to 100 million points per second if the file format matches the vertex buffer format. Since it does not need LOD structures, the progressive rendering approach can render already loaded points right away while additional points are still being loaded.

Our third contribution improves the quality of LOD-based point-cloud rendering by introducing a continuous level-of-detail approach that produces gradual transitions in point density, rather than the characteristic and noticeable blocks from discrete LOD structures. It is mainly targeted towards VR applications, where discrete levels of detail are especially noticeable and disturbing, in a large part due to the popping of chunks of points during motion.

Additional Files and Images

Weblinks

No further information available.

BibTeX

@phdthesis{SCHUETZ-2021-DISS,
  title =      "Interactive Exploration of Point Clouds",
  author =     "Markus Sch\"{u}tz",
  year =       "2021",
  abstract =   "Laser scanning, photogrammetry and other 3D scanning
               approaches generate data sets comprising millions to
               trillions of points. Modern GPUs can easily render a few
               million and up to tens of millions of points in real time,
               but data sets with hundreds of millions of points and more
               require acceleration structures to be rendered in real time.
               In this thesis, we present three contributions to the state
               of the art with the goal of improving the performance as
               well as the quality of real-time rendered point clouds.  Two
               of our contributions address the performance of LOD
               structure generation. State-of-the-art approaches achieve a
               throughput of up to around 1 million points per second,
               which requires users to wait minutes even for smaller data
               sets with a few hundred million points. Our proposed
               solutions are: A bottom-up LOD generation approach that
               creates LOD structures up to an order of magnitude faster
               than previous work, and a progressive rendering approach
               that is capable of rendering any point cloud that fits in
               GPU memory in real time, without the need to generate LOD
               structures at all. The former achieves a throughput of up to
               10 million points per second, and the latter is capable of
               loading point clouds at rates of up to 37 million points per
               second from an industry-standard point-cloud format (LAS),
               and up to 100 million points per second if the file format
               matches the vertex buffer format. Since it does not need LOD
               structures, the progressive rendering approach can render
               already loaded points right away while additional points are
               still being loaded.   Our third contribution improves the
               quality of LOD-based point-cloud rendering by introducing a
               continuous level-of-detail approach that produces gradual
               transitions in point density, rather than the characteristic
               and noticeable blocks from discrete LOD structures. It is
               mainly targeted towards VR applications, where discrete
               levels of detail are especially noticeable and disturbing,
               in a large part due to the popping of chunks of points
               during motion. ",
  month =      apr,
  address =    "Favoritenstrasse 9-11/E193-02, A-1040 Vienna, Austria",
  school =     "Research Unit of Computer Graphics, Institute of Visual
               Computing and Human-Centered Technology, Faculty of
               Informatics, TU Wien ",
  keywords =   "point cloud rendering, lidar",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2021/SCHUETZ-2021-DISS/",
}