lhoangan/arap_flow

question on para_gen.py

hoshins opened this issue · 4 comments

hoshins commented

i tried to run para_gen.py to create dataset with below arguments.

python3 para_gen.py --multseg --input /raid/hsson/datasets/DAVIS-480p --output /datasets/DMO/fd1 --fd 1 --img-pattern /datasets/FingerCamera

but there are some problems.

  1. for first some images deformed and changed background works for input (frame1) and output(frame2)
    but after that, deformation and changing background only applied to input (frame1), not for output (frame2).
lhoangan commented

Hi @hoshins , the multseg option isn't really working at this moment (my apology for not making it clear). Could you please try again without that option? If you can work it out, I'll be happy to review the pull requests.

hoshins commented

Thank you for your reply. i tried to run without multiseg option, but similar result.
as far as i think, it seems like that some deformation command skipped due to no gpu queue.
when i test with --gpu 0 1 2 3 or -gpu 0 0 0 0, it works.
in my opinion, it need to change to prevent skip.

hoshins commented

one more question.

it seems like deepmatching works but segmentation fault occured. (same result with build verification in readme.txt)
in readme file, there is metioned that not enough memory cause segmentation fault, but my environment has enough memory.
is there anything that i try to avoid segmentation fault?

hoshins commented

one more question.

when i see deformation results, there are significant artifacts on deformed images, such as bike-packing, hokey even though frame distance is just 1.
are these images having artifacts can be used for training?