Preset=1 for "enforcing 1x real-time execution" leads to failed tracking

[LenaJoachim]

Hi!

When I run the run_dso_euroc example for the EUROC dataset MH_01_easy with preset=1, the tracking fails every time (tested with two different machines). At the beginning it estimates a completely wrong trajectory and at some point it drifts away and the program finishes with:

I0303 17:37:58.151147 5455 FullSystem.cc:73] *** taking frame 1380 ***
I0303 17:37:58.152616 5455 FullSystem.cc:114] tracking new frame
W0303 17:37:58.168910 5455 FullSystem.cc:360] BIG ERROR! tracking failed entirely. Take predicted pose and hope we may somehow recover.
I0303 17:37:58.168933 5455 FullSystem.cc:378] Coarse Tracker tracked ab = -1.25229 63.7646 (exp 1 ). Res nan
W0303 17:37:58.168937 5455 FullSystem.cc:120] Initial Tracking failed: LOST!
I0303 17:37:58.168941 5455 run_dso_euroc.cc:401] Lost!
I0303 17:37:58.169009 5453 FullSystem.cc:1869] MAPPING FINISHED!
I0303 17:37:58.169126 5455 LoopClosing.h:67] wait loop closing to join
I0303 17:37:58.173430 5454 LoopClosing.cc:44] find loop closing thread need finish flag!
I0303 17:37:58.173511 5455 LoopClosing.h:82] Loop closing thread is finished
I0303 17:37:58.174085 5455 FullSystem.cc:1929] total keyframes: 3
[...]

The same happens for my own image data (12Mpix, global shutter, 5fps).

In the paper about DSO it is stated for real-time execution:

"In some rare cases - in particular during strong exposure changes - this causes initial image alignment to fail."

This failure of initial image alignment seems to be the case here, but my own images have a fixed exposure time and the lighting conditions were stable as well. Thus, strong exposure changes as a reason are quite unlikely.

What could be another reason for this behavior?

Thank you for your advice!

[NikolausDemmel]

Does it work w/o realtime?

5pfs is quite a low frame rate. So fast motions might be an issue. Just a wild guess.