Incremental SfM Loop Closure Problem in Circular Trajectories

[bxueaa]

I'm experiencing a persistent loop closure issue with COLMAP's incremental SfM reconstruction when processing a circular trajectory dataset:

1. My dataset consists of sequential frames captured along the same circular path (2 complete loops)
2. I found that COLMAP do incremental reconstruction, COLMAP begins from a certain point on the trajectory (varying based on parameters) and expands bidirectionally

3. The algorithm successfully:
a. Reconstructs most of the trajectory
b. Correctly identifies and matches overlapping regions between the two loops
c. Performs well regardless of which initial images are selected
4. Critical issue:
a. At the final stage where the two expansion directions should meet and close the loop, COLMAP consistently fails to establish matches
b. The failure position is not fixed but depends on the starting point of reconstruction
c. Visually, the images at the failing closure point appear nearly identical
d. Changing parameters alters the starting point, which shifts the closure failure to a different position on the trajectory

【Different parameters result in different starting and ending positions of the algorithm】

I think this means that the data quality is okay, and the location of this failure does not necessarily mean that the low data quality caused the loop
So is there a problem with my settings?

[bxueaa]

I can't seem to find a place to manually add matches in the GUI of Windows.
I fantasize about manually adding associations to the last place, maybe it can solve the problem.

[bxueaa]

I try to use colmap，and glomap, they all can not do well for the data.

what paramters I should set to?

[bxueaa]

i try to

[deniaqgoog-droid]

How did you solve it?

[bxueaa]

How did you solve it?

I did manage to solve it in the end, but it seems like more than a year has passed, and I've forgotten the details.

I really went through numerous parameter adjustments, changed the starting position of my data, and so on—countless operations.

But overall, I vaguely remember it took me about a week to figure it out. It was probably a combination of multiple methods that finally helped me solve the issue.