search

KRA2008 / crosscam

WYSIWYG stereoscopic 3D cross view camera mobile app
21 stars 6 forks source link

investigate whether better auto-alignment can be achieved by a second pass #127

Open KRA2008 opened 5 years ago

KRA2008 commented 5 years ago

we currently do an inverse pyramidal approach with a euclidean transform, what if after that completes we do another pyramidal but allow the other parameters to move. i have seen numerous captures in which zooming specifically would help, but zooming is not allowed in the current transform.

Goheeca commented 5 years ago

That could work well. If we allow a full-fledged homography, it wouldn't be optimal. While we want to correct a vertical (perpendicular to the mean of epipolar planes) trapezoidal distortion, we don't want to completely undo a horizontal (along the mean of epipolar lines) trapezoidal distortion. However, finding a homography and then suppressing (or better making the amount of suppression parametric) the correction of a horizontal trapezoidal distortion could be acceptable.