Why not ORB?

[liu-zg15]

Thank you for your remarkable work.
I just have a question：

I noticed LDSO detect corners by shi-tomasi.
I just wondering why don't you use ORB features like orb-slam2?
Because when calculating the descriptor, LDSO also use the ORB descriptors.

Thank you so much for your reply!

[NikolausDemmel]

Not sure why @gaoxiang12 decided this way.

My guess would be that Shi Tomasi corners are slightly more repeatable compared to FAST corners.

For FAST you can get a lot of features and it's super efficient. Since we don't use it for tracking, we neither need a lot of them, nor do we need to extract them in every frame (only keyframes), so performance and count is not of upmost importance. Repeatibility o.t.o.h. is more important for loop closures.

Otherwise, the descriptor is like ORB as you have noticed, so overall there is probably only very little difference.

[liu-zg15]

Not sure why @gaoxiang12 decided this way.

My guess would be that Shi Tomasi corners are slightly more repeatable compared to FAST corners.

For FAST you can get a lot of features and it's super efficient. Since we don't use it for tracking, we neither need a lot of them, nor do we need to extract them in every frame (only keyframes), so performance and count is not of upmost importance. Repeatibility o.t.o.h. is more important for loop closures.

Otherwise, the descriptor is like ORB as you have noticed, so overall there is probably only very little difference.

Thank you so much for your reply. That solved my question.
Another thing when I test LDSO is that:
when I test on my own sequence(with no photometric caliration), I found that the programe runs well. However, camera was stucked at some point.
When I run DSO, it runs very well and trajectory is correct.
As I know, LDSO only add a loop closing thread. The VO process is totally same with DSO(except point selection).
Do you have any idea why this appears?
Thanks again!

[NikolausDemmel]

I don't know. It's pretty much the same, but Xiang refactored the code a little bit, so there might be slight changes compare to DSO other than just the point selection.