Open jaderberg opened 11 years ago
Which dataset do you use for the swtcreate computation?
ICDAR 2003 scene test set. My laptop ran out of memory when swtcreate was using around 2.5GB.
It is expected to use a few GiBs for swtcreate, if you are not sure if this is memory related, turn off the 1GiB cache in ccv_enable_cache line in swtcreate.c
I am not sure if there is a memory leak or it is expected to run to 2GiB, all functions that is implemented in lib/*.c are checked with valgrind for memory leak issues, but swtcreate.c is not.
disabling the ccv_enable_cache line does allow it to run on my mac osx laptop for longer, however I still run out of memory and the program stops computing, and I still encounter a segmentation fault on a linux cluster
@jaderberg I will verify it with valgrind before the next release.
I am currently trying to reproduce your results for the SWT text detection performance by using bin/swtcreate.c. I run in to a segmentation fault however during the search for std_ratio. Do you have any idea what the issue could be? Also could you publish the parameters used to produce the results here http://libccv.org/doc/doc-swt/?
Update: After testing on my laptop as well as a cluster it appears there is a memory leak somewhere causing an excessive use of memory