search

alex04072000 / FuSta

Hybrid Neural Fusion for Full-frame Video Stabilization
https://alex04072000.github.io/FuSta/
550 stars 80 forks source link

To much time to process videos....even low res 320x240 and smaller. on GTX 1070 8GB #5

Open zelenooki87 opened 3 years ago

zelenooki87 commented 3 years ago

gpu usage is 95-100%... tried on 5min video 176pix ....with around 5000frames... after 10hours first step not completed.... results on your demo are great, but realy do not see point of project as low res video stabilisation takes forever. compiled code normally with miniconda, on windows 10.... if I need to change platform to ubuntu please tell me thank you

stefanklut commented 3 years ago

Not the author, but the first part of the implementation was done rather inefficiently. Where no matter how small/big the input is the images will always be resized to 832 by 448. Considering this is 70% of the process time and it takes 9 seconds per image on images of this size, 10 hours is to be expected with 5000 frames.

zelenooki87 commented 3 years ago

thank you for explanation. may I ask what is your suggestion to try for video stabilisation, using deep learning open source code here on github? until developer of fusta reorganize and optimize code.

alex04072000 commented 3 years ago

@stefanklut Thank you for your kind explanation. @zelenooki87 As said by @stefanklut , the first part of the code (flow smoothing) is not implemented by us. We use this implementation as this is the best-performing method of flow smoothing. A quick solution is to reduce the running time is to trim down the frame number (e.g. using the first 500 frames) or skip some frames periodically (e.g. only process the odd index of frames).

stefanklut commented 3 years ago

If you don't mind cropping maybe try https://github.com/Annbless/DUTCode. But if you don't want cropping right now this is your best bet, considering it's much less distortion compared to DIFRINT.