Real-time Meshing for Noisy Points

Daniel Prieler
Real-time Meshing for Noisy Points
[poster] [thesis]

Information

Abstract

The increasing availability of 3D scanning devices in both industrial and entertainment environments (e.g., Microsoft Kinect) creates a demand for fast and reliable resampling and reconstruction techniques. Point clouds, especially raw range images, are often non-uniformly sampled and subject to non-uniform noise levels. Current state-of-the-art techniques often require user-provided parameters that estimate the noise level of the point cloud. This produces sub-optimal results for point sets with varying noise extent. We propose an isotropically fair neighborhood definition which is specifically designed to address non-uniformly sampled point clouds. Our iterative point cloud resampling method estimates and adapts to the local noise level at each sample. This increases the reconstruction quality for point clouds with high noise levels while being completely parameter free. The data structures built during the resampling process are reused to speed up the process of creating a consistent normal orientation. Evaluation of the re- sampling quality shows that our technique outperforms current state-of-the-art methods for varying noise levels and non-uniform sampling. Both the resampling algorithm and the subsequent consistent normal orientation operate locally and can be implemented efficiently in parallel. Our GPU sphere regression implementation outperforms the stan- dard sequential procedure by a factor of 20.

Additional Files and Images

Additional images and videos

image: Local orienting of normals image: Local orienting of normals

Additional files

Weblinks

No further information available.

BibTeX

@mastersthesis{prieler_daniel-2013-da,
  title =      "Real-time Meshing for Noisy Points",
  author =     "Daniel Prieler",
  year =       "2016",
  abstract =   "The increasing availability of 3D scanning devices in both
               industrial and entertainment environments (e.g., Microsoft
               Kinect) creates a demand for fast and reliable resampling
               and reconstruction techniques. Point clouds, especially raw
               range images, are often non-uniformly sampled and subject to
               non-uniform noise levels. Current state-of-the-art
               techniques often require user-provided parameters that
               estimate the noise level of the point cloud. This produces
               sub-optimal results for point sets with varying noise
               extent. We propose an isotropically fair neighborhood
               definition which is specifically designed to address
               non-uniformly sampled point clouds. Our iterative point
               cloud resampling method estimates and adapts to the local
               noise level at each sample. This increases the
               reconstruction quality for point clouds with high noise
               levels while being completely parameter free. The data
               structures built during the resampling process are reused to
               speed up the process of creating a consistent normal
               orientation. Evaluation of the re- sampling quality shows
               that our technique outperforms current state-of-the-art
               methods for varying noise levels and non-uniform sampling.
               Both the resampling algorithm and the subsequent consistent
               normal orientation operate locally and can be implemented
               efficiently in parallel. Our GPU sphere regression
               implementation outperforms the stan- dard sequential
               procedure by a factor of 20.",
  month =      feb,
  address =    "Favoritenstrasse 9-11/186, A-1040 Vienna, Austria",
  school =     "Institute of Computer Graphics and Algorithms, Vienna
               University of Technology",
  keywords =   "surface fitting, surface reconstruction, noise, meshing,
               real-time, CUDA",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2016/prieler_daniel-2013-da/",
}