search

diegoaldarondo / Label3D

Labeling GUI for multi-camera tracking
33 stars 11 forks source link

Labeling on the undistort images/video #18

Open EvenGu opened 2 years ago

EvenGu commented 2 years ago

Hi, I could be wrong but I wonder if it would be more accurate if we do the labeling on the undistort images? By undistort images, I mean using the camera parameters we get from calibration. Also I think that in DANNCE, the Lframe calibration is based on undistort images. Looking forward to your reply. Thanks!

diegoaldarondo commented 2 years ago

Hi Even,

The default setting assumes that the images are distorted. I believe I included support for labeling undistorted images by setting the undistortedImages parameter to true.

The relevant portions of the program that treat the two conditions differently are https://github.com/diegoaldarondo/Label3D/blob/f880cc2576d7442fd242336844c1d9095ae7f056/Label3D.m#L652 and https://github.com/diegoaldarondo/Label3D/blob/f880cc2576d7442fd242336844c1d9095ae7f056/Label3D.m#L735. The program transforms the labeled points into the (undistorted) calibration coordinates if needed and reprojects them into the distorted coordinates if needed.

Either way, in my experience it hasn't made a difference with good intrinsic calibration. I haven't played much with cameras with high lens distortion, so it would be interesting if it started to matter there. Feel free to report back any results here.

Best, Diego

EvenGu commented 2 years ago

Hi Diego, Thanks for replying. I didn't see much improvement in using undistort images either. The reason that I am thinking about undistorted images is that my triangulation is sometimes off in some views; my reprojected 2D points are not always correct...Do you have any suggestions on that?

diegoaldarondo commented 2 years ago

This is probably a sign that the extrinsic calibration is not correct. I would recommend recalibrating and trying again.