Michael Eickmeyer
Tile Coding for Directional Path Guiding

Information

Abstract

Path guiding is a promising Monte Carlo ray tracing technique in physically based rendering (PBR) that leverages information about the scene to improve convergence and reduce image noise. By employing data structures that approximate the incident radiance, such algorithms are capable of making informed decisions about where to send future rays. However, even though minimizing noise is a major goal in path guiding, current state-of-the-art techniques barely take the underlying sample noise into account when building guiding distributions, solely relying on the illumination itself. To address this issue, this thesis proposes two novel data structures based on tile coding, a discrete data structure that has seen success in reinforcement learning due to its high degree of generalization. By making use of its advantages in combination with a data-driven approach to quantify sample noise, the data structures approximate the incident radiance with a higher level of detail, while minimizing typical issues other discrete data structures exhibit. To prove their effectiveness, the data structures are compared on a dataset of 446 environment maps against discrete and continuous state-of-the-art approaches, outperforming them in their ability to represent the incident radiance field in the majority of equal-sample comparisons. As part of a larger project, this work provides the directional data structure as the foundation for an upcoming complete path guiding algorithm.

Additional Files and Images

No additional files or images.

Weblinks

No further information available.

BibTeX

@bachelorsthesis{eickmeyer-2027-tcd,
  title =      "Tile Coding for Directional Path Guiding",
  author =     "Michael Eickmeyer",
  year =       "2027",
  abstract =   "Path guiding is a promising Monte Carlo ray tracing
               technique in physically based rendering (PBR) that leverages
               information about the scene to improve convergence and
               reduce image noise. By employing data structures that
               approximate the incident radiance, such algorithms are
               capable of making informed decisions about where to send
               future rays. However, even though minimizing noise is a
               major goal in path guiding, current state-of-the-art
               techniques barely take the underlying sample noise into
               account when building guiding distributions, solely relying
               on the illumination itself. To address this issue, this
               thesis proposes two novel data structures based on tile
               coding, a discrete data structure that has seen success in
               reinforcement learning due to its high degree of
               generalization. By making use of its advantages in
               combination with a data-driven approach to quantify sample
               noise, the data structures approximate the incident radiance
               with a higher level of detail, while minimizing typical
               issues other discrete data structures exhibit. To prove
               their effectiveness, the data structures are compared on a
               dataset of 446 environment maps against discrete and
               continuous state-of-the-art approaches, outperforming them
               in their ability to represent the incident radiance field in
               the majority of equal-sample comparisons. As part of a
               larger project, this work provides the directional data
               structure as the foundation for an upcoming complete path
               guiding algorithm.",
  month =      jan,
  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 =   "physically based rendering, Monte Carlo rendering, path
               tracing, importance sampling, path guiding, data structures",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2027/eickmeyer-2027-tcd/",
}