Shape from Motion

Search IconIcon to open search

Initial paper?: https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7010934https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7010934

Goal

reconstruct the surface of an object

  • reference 3D shape (template) of the object is available
  • under a specific deformation constraint

Source: lamarca-2020

SFT (shape from template)

  • uses only a single image — faster than nrsfm
  • lower computational cost
  • must have a known 3D template (textured model)

SfT methods

  • require:
    • 1 monocular image
    • 1 textured shape at rest (template) “geometry” as the deformation model
  • different definitions of the deformation model
    • analytic, e.g. isometric deformation-based: assumes preserved geodesic distance between surface points
      • isometry for SfT has proven to be well-posed –> led to stable, real-time solutions
    • energy-based; jointly minimises {energy shape w.r.t. template [shape at rest] + reprojection error for image correspondences}
  • classification according to [ http://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Gallardo_Shape-From-Template_in_Flatland_2015_CVPR_paper.pdf ]
    • statistics-based
      • use data to learn the space of deformations
      • effective in low-dimensional spaces
    • physics-based
      • uses mathematical models (based on physical laws) in order to compute the space of deformations
      • e.g. isometric model