search

rpng / open_vins

An open source platform for visual-inertial navigation research.
https://docs.openvins.com
GNU General Public License v3.0
2.13k stars 630 forks source link

use bearing vector for feature coordinate representation instead of normalized uv #218

Closed CanCanZeng closed 2 years ago

CanCanZeng commented 2 years ago

Hi, would you please use bearing vector for feature coordinate representation instead of normalized uv? My suggestion is based on the thought that, the bearing vector is more general for all kinds of camera models, including pinhole or fisheye or omnidirectional camera. As far as I know there are two parts need to change. One is triangulation, and the other is mis-matches invalidation in feature tracking. The triangulate function does not need to change a lot, the bigger change happens to invalidate mis-matches in feature tracking, we cannot use cv::findFundamentalMat but need to realize a customized findEssentialMat. For this part we can refer to openvslam https://github.com/OpenVSLAM-Community/openvslam/blob/cf452c3dbc31b5a1346a14912fbdf14aa42752c7/src/openvslam/solve/essential_solver.cc#L89

I can help if you would like to do so, since I have done this to vins-mono before.

goldbattle commented 2 years ago

I don't think this is something we want to have included currently, but feel free to make a fork with support if you want others to have this as an option. We already support pinhole and fisheye so the only reason is for > 150 degree cameras. This is already tracked in issue #184 I believe. Additionally, support for calibration and documentation of these cameras would need to be added which isn't trivial.

CanCanZeng commented 2 years ago

You are right, my suggestion is a solution for #184 .