Downscale the extracted frames from video (n3d)

[remmel]

Add downscale parameter to resize the frames right after their extractions in order to save RAM and hard disk (and GPU VRAM?) Doing that allow me to run flame_steak on my laptop with RTX3060-mobile 6GB VRAM and 20GB RAM

Remind that there is also the resolution param in config. Thus if we have --downscale 2 and config/resolution: 4 final image takes into account will be 2*4=8

python script/pre_n3d.py --videopath flame_steak --downscale 4 --startframe 200 --endframe 225

I also rework extractframes fn to avoid extracting useless frames (save time and hard disk)

[remmel]

Thank you for your work ! Could you keep the original pre_n3d and rename your pre_n3d to other name, like pre_n3d_with_scale ? pre_n3d is for reproducing purpose. resolution and number of frames are fixed in the pre_script.

Hi Li, My final objective is to avoid having duplicated code, thus I can't duplicated pre_n3d file (to have pre_n3d_with_scale). I understand that you don't want to change the default behaviour, however in my case if the downscale is not indicated it will be 1 (thus no change). I agree that now by default, if endframe is not indicated, extractframes will extract 50 frames (parser.add_argument("--endframe", default=50, type=int)) instead of 300; as only thoses 50 frames are used in the reconstruction. No worries, I can discard my PR, that pre downscale feature is not no usefull when you have a good GPU; or update the code to always extract 300 frames; or whatever you think.

[lizhan17]

Thank you for your work ! Could you keep the original pre_n3d and rename your pre_n3d to other name, like pre_n3d_with_scale ? pre_n3d is for reproducing purpose. resolution and number of frames are fixed in the pre_script.

Hi Li, My final objective is to avoid having duplicated code, thus I can't duplicated pre_n3d file (to have pre_n3d_with_scale). I understand that you don't want to change the default behaviour, however in my case if the downscale is not indicated it will be 1 (thus no change). I agree that now by default, if endframe is not indicated, extractframes will extract 50 frames (parser.add_argument("--endframe", default=50, type=int)) instead of 300; as only thoses 50 frames are used in the reconstruction. No worries, I can discard my PR, that pre downscale feature is not no usefull when you have a good GPU; or update the code to always extract 300 frames; or whatever you think.

hi remmel,

For each scene in n3d dataset, 6 50-frame-length sub-sequences are reconstructed. So our code extracts 300 frames during preprocess. That is just for reproducing purpose. users wont't need to re-run the preprocess code if they need 300 frames.

Can you update the code to always extract 300 frames, or 300 frames by default, then, i can merge the code.

I like the use of pathlib in the preprocess part. Thank you for your work.

[remmel]

Hi Li,

Understood; I've just updated the code, to always extract 300 frames like before.