Author: Hicham Hadj-Abdelkader
Year: 2008
- perspective image tools are unable to handle geometrical distortion
- images are mapped onto a unit sphere and treated in the spherical spectral domain
- feature based: extracting features + matching and pose estimation (low accuracy but wide baselinge)
- template based: using the whole images (high accuracy but small baseline)
- Here uses feature based method to initialize template based method
- focus on feature based using Harris point detector and SIFT descriptor
- mapping of the image on the unit sphere:
- sampling a equi angular grid on the unit sphere
- spherical point mapped on the image plane with camera model
- interpolation in the omnidirectionnal plane
- spherical spectral domain:
$I$ can be expanded into a linear summation of spherical harmonic function - convolution is defined in
$SO(3)$ for spherical images
- gaussian filtering is usually performed before feature extraction => spherical gaussian filter
- Definition of spherical derivatives (w.r.t.
$\theta$ and$\phi$ )
- use spherical derivatives for Harris corners
- comparison using the repeatability rate
- SIFT descriptors are built on a circular region in the spherical image
- Odometry using only
$E$ estimation with RANSAC - => improvement vs classical methods