The Meshroom 2020.1.0 crashes when Akaze is enabled. The error usually occurs in the FeatureExtraction step. A job with a small quantity of images (like 32) seems to do okay, but larger image numbers increase the likelihood of failure. Sift alone seems to work fine.
I spent a lot of time testing the possibility that a freshly booted system may have uninitialized hardware. I would let my first try crash and then delete the cache. I would then restart my saved .mg file and things seemed to run better. My 144 image job would now finish FeatureExtraction, when it was crashing earlier. I then attempted my 2300 image job and it still would crash at some chunk in FeatureExtraction.
This morning, I have physically removed one of my GPUs (a GTX Titan Black) and rebooted, to test the system with only one GPU. The 2300 image job crashed in chunk '2' of the FeatureExtraction step. I deleted the cache and restarted Meshroom. I ran the same 2300 image job (with only the Quadro P2000 gpu in the system) and this time the program ran further and finally failed at chunk #10. For some reason, the program seems to be able to run further after deleting a failed cache and restarting.
This problem happens regardless of the 'force cpu' setting, in FeatureExtraction. A similar setup, using the 2300 image source and Akaze runs without error under Linux.
Any attempt to restart Meshroom and start the same job with a failed cache displays the message “Graph is being computed externally”. What graph? The one displayed in the Statistics window? Is that a symptom or cause of the problem? By the way, the Statistics tab will only display one GPU, even if two are in the system.
I am conducting these tests, while monitoring my system resources and do not see any resource being exhausted. I am ready to test anything I can. Please guide me in helping resolve this problem.