Direct Fisheye Stereo Correspondence Using Enhanced Unified Camera Model and Semi-Global Matching Algorithm
Author: Khomutenko
Year: 2018
- Semi-Global matching alg well known for stereo corespondence
- omnidirectionnal LSD SLAM => curve sampling is way more computationnally expensive than searching on lines
Enhanced Unified Model:
- It is based on a projection surface (of revolution): the image point is obtained by scaling the 3D point on the surface and then projecting it orthogonally on the image plane (cf fig 1)
- For pinhole model, the projection surface is the plane z=1
- With the constraints of a projection surface, we can get the projection surface equation from the parameters of the model
- Then straight lines are projected as conic sections (the only straight lines projected as straight lines are those going through the optical center)
- Develop the equation of the conic section of a given line (that defines a plane with the optical axis) these are epipolar curves
Stereo matching:
- The stereo equation with essential matrix gives the equation of the plane => gives the conic equation
- we get a polynomial with a dirty equation :/
- we need to rasterize the curve ie convert it into pixels (this is the expensive part)
- rasterization: start from a point (the epipole) and follow the curve thanks to the gradient until another point (the other epipole)
Semi Global matching algorithm:
- Require a table of photometric error for all pixels and disparity (if the disparity is evaluated on 64 level we need width x height x 64 of memory) here they use blocks of 3 x 3 pixels to reduce the memory needed
- For every epipolar curve, a narrow band is computed around it and the average photometric error is computed on each block of the band
- Dynamic programming: find a path in the error buffer that is optimal
- programmation dynamique = trouver une solution optimale en combinant des solutions optimales à des sous problèmes