simonmeister / UnFlow

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

Questions about image for loss comparisons and for neural network #44

Closed Blcony closed 5 years ago

Blcony commented 6 years ago

Hi, simon! Thank you for your code and work. Now I have some questions, I hope you can give me some advices. In the file unsupervised.py, line 68 says that images for neural network input with mean-zero values in [-1, 1]. But in common sense, the images for neural network should be subtracted by their mean, and divided by their stddev. Why do you just divide images by 255.0? Why should the images for network have the value in [-1, 1]? And line 65 says that images for loss comparison with values in [0,1], I wonder whether this operation is necessary? Can I directly set the raw images with values in [0, 255.0] to calculate the loss? Thank you very much!