Maybe it's me and maybe I am doing something wrong, but I feel like the results I am getting out of Photoscan, while being achieved in shorter time, are actually getting worse in quality.

I recently tried to produce a reconstruction of a large image set of the STRI BCI permanent plot.  This set contains about 3300 photos.  the main body of pictures (4 pairs of flight lines) should form a nice normal rectangle (about 600 m x 1100), but instead there is a pronounced 'splaying' of the flight lines.   These flights were carried out with very high side overlap (75%) which I have confirmed based on manually registering sets of frames from parallel tracks.  In addition, this Photoscan run was carried out using the 'Ground Control' method in which estimates of camera location (XYZ) and provided by the user to improve to reconstruction (in this case I derived those estimates from our  UAV GPS, see Stephen's blog here).

However, after reconstruction, the cameras that are in the greatest 'splayed' areas to the west and left of the dataset have huge location errors greater than 150 m! It seems like what is happening here is that Photoscan is not identifying any 'cross-track' matches, in other words, it does not know that parallel images might be related to each other and worth evaluating matches.  I ran a test of just a handful of photos that are in parallel tracks and the reconstruction seems to be a success.  I wonder if part of their optimization is to only work on matching images 'along-track' or only trying to conduct matches on the images that are in front of or behind a given image.  These results were also obtained when images were matched in 'generic' pair-preselection mode.  The final option is to 'disable' generic pair preselection, which I believe means that Photoscan has to try to compare every image to every other image, taking substantially more time, a model which is similar to that of Bundler.

I plan to run some tests with other settings and will soon be putting it all into Ecosynther to see how that performs.

The Photoscan forums were down at the time of my writing this, so I haven't been able to check out my results there.    Just wanted to get my notes out there and see what others are thinking.  I have tried using the Photoscan 'optimize' magic wand to improve the camera location results in Ground Control, but that too didn't help.  

Views: 437

Reply to This


© 2019   Created by Erle Ellis. Content is made available under CC BY 4.0.   Powered by

Badges  |  Report an Issue  |  Terms of Service