Improved 3-D Scene Sampling By Camera Model Design

Images are one of the most powerful means of communication available. They are pervasive throughout our lives and are the central focus of computer graphics, visualization, and computer vision. Most images, synthetic or photographic, are created using the planar pinhole camera model. This classic camera model has important advantages including its simplicity, enabling efficient hardware and software implementations, and its similarity to human vision, yielding images familiar to users. However, the planar pinhole camera model suffers from important limitations including sampling from a single viewpoint and requiring a uniform sampling rate along the image plane. These limitations result in problems with occlusions, when no direct line-of-sight exists to the viewpoint, and sampling rates which do not correlate well to the complexity of 3-D data. This dissertation proposes a new paradigm of problem solving, dubbed Camera Model Design, which overcomes the limitations of the planar pinhole camera model to address many problems which still exist in computer graphics, visualization, and computer vision. The Camera Model Design paradigm stresses four important ideas. First, relax the constraints of the planar pinhole camera model allowing generalized camera rays which are no longer straight and no longer converge. This facilitates camera models that overcome occlusions and have variable sampling rates. Second, the choice of camera used for a particular application need not be limited to the planar pinhole camera. Instead, a new camera should be designed to directly address the needs of the application. Third, camera models should no longer be static. Instead they should dynamically adapt to the 3-D data they are sampling. Finally, in order to support interactive exploration, a high level of computational efficiency should be maintained. We introduce three new families of camera model, the occlusion cameras, the graph cameras, and the general pinhole cameras, all of which address one or more of the planar pinhole camera limitations. These new camera models are applied to a wide variety of applications which demonstrate the benefits of Camera Model Design for increasing computational efficiency, improving output image quality, and enhancing user performance in exploring 3-D datasets.

Paul Rosen. Improved 3-D Scene Sampling By Camera Model Design. Purdue University, 2010.

@phdthesis{rosen2010improved,
  title = {Improved 3-D scene sampling by camera model design},
  author = {Rosen, Paul},
  school = {Purdue University},
  year = {2010},
  keywords = {Computer graphics, Computer vision, Human interfaces, Multiperspective 
images, Occlusion, Three-D scene sampling, Camera},
  abstract = {Images are one of the most powerful means of communication available. They
    are pervasive throughout our lives and are the central focus of computer graphics,
    visualization, and computer vision. Most images, synthetic or photographic, are created
    using the planar pinhole camera model. This classic camera model has important
    advantages including its simplicity, enabling efficient hardware and software
    implementations, and its similarity to human vision, yielding images familiar to users.
    However, the planar pinhole camera model suffers from important limitations including
    sampling from a single viewpoint and requiring a uniform sampling rate along the image
    plane. These limitations result in problems with occlusions, when no direct
    line-of-sight exists to the viewpoint, and sampling rates which do not correlate well to
    the complexity of 3-D data.  This dissertation proposes a new paradigm of problem
    solving, dubbed Camera Model Design, which overcomes the limitations of the planar
    pinhole camera model to address many problems which still exist in computer graphics,
    visualization, and computer vision. The Camera Model Design paradigm stresses four
    important ideas. First, relax the constraints of the planar pinhole camera model
    allowing generalized camera rays which are no longer straight and no longer converge.
    This facilitates camera models that overcome occlusions and have variable sampling
    rates. Second, the choice of camera used for a particular application need not be
    limited to the planar pinhole camera. Instead, a new camera should be designed to
    directly address the needs of the application. Third, camera models should no longer be
    static. Instead they should dynamically adapt to the 3-D data they are sampling.
    Finally, in order to support interactive exploration, a high level of computational
    efficiency should be maintained. We introduce three new families of camera model, the
    occlusion cameras, the graph cameras, and the general pinhole cameras, all of which
    address one or more of the planar pinhole camera limitations. These new camera models
    are applied to a wide variety of applications which demonstrate the benefits of Camera
    Model Design for increasing computational efficiency, improving output image quality,
    and enhancing user performance in exploring 3-D datasets.}
}