could you release the code to extract optical flow?

google-deepmind / kinetics-i3d

Convolutional neural network model for video classification trained on the Kinetics dataset.

Apache License 2.0

1.74k stars 461 forks source link

could you release the code to extract optical flow? #24

Open hjeun opened 6 years ago

hjeun commented 6 years ago

do I need to normalize RGB images to [0,1] from [0,255] before calculate optical flow? I will follow the code here. https://github.com/opencv/opencv/blob/master/samples/gpu/optical_flow.cpp And 'd_flow' is only output of TVL1. how to split dx_flow and dy_flow?

joaoluiscarreira commented 6 years ago

Hi,

from the person who wrapped TVL-1 at google: (...) "we truncate large motions at +/- 20 pixels along each dimension and then quantize values to 256 buckets using a linear map from [-20, 20] to [0, 255]."

jahab commented 4 years ago

convert_to_flow.ipynb.zip Hi, this is the .ipynb file to extract the flows from the images. This file extract images from videos and then convert frames to optical flow files.