First of all, thank you for this wonderful software!
It looks like the global BA pipeline uses ReprojErrorCostFunctor, which computes residuals in pixel units.
But when filtering outliers in GlobalMapper::GlobalPositioning(), the points are filtered based on the normalized reprojection distances (distances on the normalized image plane).
I'd like to ask why does ReprojErrorCostFunctor compute the residual in pixel units instead of using the normalized reprojection error like the point-filtering method does?
If there are 2 cameras, CamA with the resolution 640x480 px, and CamB with 3840x2160 px, then reprojection error of 1 px is more severe for CamA than for CamB. Doesn't the lack of normalized reprojection error introduce bias into the optimization?
The distortion is also computed in the normalized image plane, so why not the reprojection error too?
I'm sure you had a very good reason to choose not to normalize the residual in ReprojErrorCostFunctor, but I can't figure out why is it better than using the normalized error. In my mind, the normalized error would make the error independent of cameras' resolution, which would remove bias.
Thanks!
First of all, thank you for this wonderful software!
It looks like the global BA pipeline uses
ReprojErrorCostFunctor, which computes residuals in pixel units.But when filtering outliers in
GlobalMapper::GlobalPositioning(), the points are filtered based on the normalized reprojection distances (distances on the normalized image plane).I'd like to ask why does
ReprojErrorCostFunctorcompute the residual in pixel units instead of using the normalized reprojection error like the point-filtering method does?If there are 2 cameras,
CamAwith the resolution 640x480 px, andCamBwith 3840x2160 px, then reprojection error of 1 px is more severe forCamAthan forCamB. Doesn't the lack of normalized reprojection error introduce bias into the optimization?The distortion is also computed in the normalized image plane, so why not the reprojection error too?
I'm sure you had a very good reason to choose not to normalize the residual in
ReprojErrorCostFunctor, but I can't figure out why is it better than using the normalized error. In my mind, the normalized error would make the error independent of cameras' resolution, which would remove bias.Thanks!