bundle_adjustment does not work well

74 views
Skip to first unread message

l pp

unread,
Jul 12, 2022, 5:02:43 AM7/12/22
to Ames Stereo Pipeline Support
hello,
I am trying to generate a high quality DEM using the Resource 3 satellite image. As suggested in the book, I tried to use Bundle Adjustment first, then parallel_STEREO, and finally generate DEM. However, I found that the result of using bundle adjustment was worse than that of not using bundle adjustment. I don't know why.The ASP version is 3.1.1-alpha

command:
1、bundle_adjust -t rpc zy3-BWD.tiff zy3-FWD.tiff zy3-BWD.xml zy3-FWD.xml -o run_ba/run
2、parallel_stereo -t rpc --alignment-method local_epipolar --stereo-algorithm asp_mgm  
--subpixel-mode 2 --corr-memory-limit-mb 2048 zy3-BWD.tiff zy3-FWD.tiff 
zy3-BWD.xml zy3-FWD.xml run_adjust/run --bundle-adjust-prefix run_ba/run
3、point2dem run-PC.tif -o run_adjust/run --search-radius-factor 5 --dem-hole-fill-len 500 
--nodata-value -9999
DEM results:
1.jpg2.jpg

Alexandrov, Oleg (ARC-TI)[KBR Wyle Services, LLC]

unread,
Jul 12, 2022, 12:21:04 PM7/12/22
to l pp, Ames Stereo Pipeline Support
Did you change anything except using or not using the "--bundle-adjust-prefix" option? So, except using or not using bundle-adjusted cameras? 

I will suggest running point2dem --errorimage on both point clouds, with and without bundle adjustment, then inspecting the obtained intersection error image. One can use gdalinfo -stats to see their stats, then colorize it with colormap with the same --min and --max values in both cases, then open that in stereo_gui to see how much difference the bundle adjustment makes. It is meant to reduce the intersection error. 

You can also try "--alignment-method affineepipolar".  Then you can try creating a lower-resolution DEM, mapprojecting onto it, and redoing stereo (if you use --bundle-adjust-prefix, then it either be used at all stages or not at all). There are more details here: https://stereopipeline.readthedocs.io/en/latest/next_steps.html#mapproj-example

Hopefully some of these will help with those blunders.  Especially mapprojection is strongly suggested for steep terrains like what you have.



From: ames-stereo-pi...@googlegroups.com <ames-stereo-pi...@googlegroups.com> on behalf of l pp <pplau...@gmail.com>
Sent: Tuesday, July 12, 2022 2:02 AM
To: Ames Stereo Pipeline Support <ames-stereo-pi...@googlegroups.com>
Subject: [EXTERNAL] bundle_adjustment does not work well
 
--
You received this message because you are subscribed to the Google Groups "Ames Stereo Pipeline Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ames-stereo-pipeline...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ames-stereo-pipeline-support/9d3a995e-56eb-4fbe-bad1-dfe53fa2350an%40googlegroups.com.

l pp

unread,
Jul 12, 2022, 10:25:41 PM7/12/22
to Ames Stereo Pipeline Support

Thank you for your advice. I will try what you suggested

l pp

unread,
Jul 19, 2022, 8:29:23 AM7/19/22
to Ames Stereo Pipeline Support
Thanks for your suggestion.
After trying to generate a low-resolution DEM first, then use mapproject, and then do stereo, the holes in the DEM disappeared. 
But I found that the resolution of DEM is relatively low and the border is blurred, so may I ask you another question, is there any suggestion to improve the resolution of DEM?
DEM.JPG

Oleg Alexandrov

unread,
Jul 19, 2022, 12:09:30 PM7/19/22
to l pp, Ames Stereo Pipeline Support
Our best algorithm is MGM, to be invoked as: parallel_stereo --stereo-algorithm mgm. 

I hope you mapprojected the left and right image at the same resolution, that can make a difference. And one should be careful with what that resolution is. That is discussed here. https://stereopipeline.readthedocs.io/en/latest/next_steps.html#resolution-of-mapprojection

You can also try to hillshade the DEM from stereo_gui, it will be easier to inspect that way. 

l pp

unread,
Jul 20, 2022, 2:41:28 AM7/20/22
to Ames Stereo Pipeline Support
Thank you for your quick reply .
Reply all
Reply to author
Forward
0 new messages