I have a forested plot with pictures using RTK that is being reconstructed as several non-overlapping point-clouds. I have some fiducial markers in the images that could connect them, but they aren’t being connected correctly. The cameras have the correct locations but if the path doesn’t have enough photos the orientation (pitch) is lost. This is leading to some areas of trees being skewed (not straight up and down). I know image locations can be manually connected using GCPs, but I don’t have the exact location of markers. Is there a way to force matching of manually tagged markers so that the cameras are oriented correctly?
It only took about an hour to process 794 photos with 4656x3492 resolution, so I would love your suggestions on increasing the settings. I have an NVidia 1080 Ti with 11GB Ram which never goes over 3GB used and 30% GPU use on sift ‘high’, but when I change settings for sift to ‘ultra’ webodm automatically reverts to CPU.
Current settings: gps-accuracy: 2, use-3dmesh: true, use-hybrid-bundle-adjustment: true
All other settings are default.
I should also mention the photos are taken from a terrestrial camera (not a UAV), all with no roll and with pitch of 0 degrees (aiming towards the horizon) but sometimes pointing up or down slightly.
Here is a typical photo:
Here is the point cloud, where you can see some sections were skewed. Note that photos were tagged with RTK GPS generally with <10cm accuracy but at worst <50cm.
The point cloud with the higher settings is beautiful, but still not aligned. I went back and took more photos, but may also try some manual control points with opensfm.
Set your gps-accuracy to .05 or even 0.01 (don’t set below 0.02 if it was more than one image capture session). This will constrain the data to the known positional accuracy, and may help significantly.
There may be some other tricks worthy of trying, but start there.
