Bug when fix_existing_images=True leading to crash #3158
Conversation
|
Thank you for this PR! Wouldn't we instead expect existing images to not be filtered out? @ahojnnes |
@sarlinpe @ahojnnes One scenario (edge case?) that could happen is when the original input reconstruction is a triangulated model from known camera poses, some images/poses don't have any triangulated 3D points but they are still in the reconstruction. Therefore, after continuing the mapping process, those images will/could get filtered out (based on some thresholds somewhere in the mapping process?), leading to this error. This is what happened to me that prompted me to make this PR, and afaik, this is the exact same scenario as #2888 (the input reconstruction is a triangulated model). |
|
Need to look at this again, sorry for the lack of response. We now have a second person confirming the issue. This patch seems to fix their issue as well. @sarlinpe Probably we should fix this before a 3.12.0 release. |
|
Does the following change fix the issue? #3403 |
When
--Mapper.fix_existing_images=True, the current code transforms the final reconstruction back to the original coordinate frame by storingorig_fixed_image_positions. Anout_of_rangebug could occur when the mapping process could filter out some of the original registered images, therefore those filtered image ids will not exist anymore in the reconstruction.As a temporary fix, I simply check again if the image id exists in the final reconstruction and build a
filtered_orig_positionsarray (this will be a subset oforig_fixed_image_positions).Should fix #2888 (see specifically the last 2 comments).