Author: Zhang
Year: 2016
- impact of fisheye or catadioptric on the quality of motion estimate of urban and indoor scene
- experiment checking the keypoint tracking accuracy (in pixel) wrt number of frame of the trajectory -> less accuracy with big FOV
- when the FOV gets bigger, the angular resolution gets smaller
- experiments solving the absolute pose estimation (pnp) problem with noisy landmarks
- experiments with a full VO pipeline: large FOV are better in confined environnement
SVO for omni cam:
- to change: implementing camera models + reprojection error with bearing vectors + sampling on epipolar curve
- choose of an error metric for bearing vector errors -> bearing vector error is the best
- sample pixel on the epipolar line => sample pixels between
$f_{min}, f_{max}$ the bearing vectors representing the depth interval we search in