search

simonmeister / UnFlow

UnFlow: Unsupervised Learning of Optical Flow with a Bidirectional Census Loss
MIT License
294 stars 57 forks source link

regarding image normalization #20

Closed stalin18 closed 6 years ago

stalin18 commented 6 years ago

Hi, I see that you perform mean normalization with:

mean = [104.920005, 110.1753, 114.785955] stddev = 1 / 0.0039216

Are these values from ImageNet dataset? I'm curious to know if performing normalization considerably affected (improved) your results? I believe none of the papers (your's included) explicitly mention about input image normalization for optical flow. Thanks!

simonmeister commented 6 years ago

The stddev is about 255, so it's simply re-scaling the range to be approximately contained in [-1, 1]. I believe for the mean values, just using 255/2 should work equally well, but i only tried training with this normalization. I used the mean from the code at "http://scs.ryerson.ca/~jjyu/projects/unsupervised/optical/flow/machine/learning/2016/08/30/Unsup-flow.html" and i think it's the KITTI mean.

wpfhtl commented 6 years ago

what about the mean of sintel and synthia. It looks you didn't use mean in synthia, but use the same mean in sintel with kitti.

simonmeister commented 6 years ago

The same mean is used everywhere.