Skip to content

Latest commit

 

History

History
28 lines (20 loc) · 1.4 KB

real_time_direct_dense_plane_sweeping.md

File metadata and controls

28 lines (20 loc) · 1.4 KB

Real-Time Direct Dense Matching on Fisheye Images Using Plane-Sweeping Stereo

Author: Häne

Year: 2014

Notes:

  • adaptation of plane sweeping (R.T. Collins - 96) alg on fisheye images
  • requirement: projection and unprojection done efficiently
  • in comparison to the rectification approach, plane sweeping can be done on multiple images
  • suitable for GPU implementations
  • Use of unified projection model

Stereo Matching on fisheye images

  • Camera projection model $[u,v] = K D( h( \xi, X))$, $D$ to take distortions into account and $\xi$ is a scalar parameter that models the fisheye length
  • Use a reference image $I_{ref}$ and a set of plane hypotheses $\Pi$ and each plane $\Pi_m$ defined by $[\mathbf{n}_m^T, d_m]$
  • the planes are used as local reconstruction of the image => for each pixel of $I_{ref}$, the best plane is chosen
  • For each plane hypothesis $\Pi_m$, the whole image $I_n$ is projected on the plane through a warped image $I_m$ and an image dissimilarity (ZNCC) is computed between each warped image $I_m$ and $I_{ref}$
  • An aggregated cost is computed to take into account several views and motion
  • Finally winner takes all strategy: the plane $\Pi_m$ with the lowest aggregated score is selected for each pixel
  • subpixel interpolation for smoothness between neighbouring planes

Commentaires

  • Comment sont choisis les plans? Ne marche qu'en envirronnement structuré mais pourquoi pas