search

simonmeister / UnFlow

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

Questions about random affine augmentation #13

Closed bryanyzhu closed 6 years ago

bryanyzhu commented 6 years ago

Hi Simon, I got another question about the random affine augmentation during training. The code snippet is here. I don't understand why there is a local augmentation on the im2_geo again. Is it because we want more distortions? I saw similar operation in Jason's UnsupFlowNet caffe implementation before. I didn't figure out either. So could you give me some help here? Thank you.

simonmeister commented 6 years ago

The two images are first geometrically distorted together as a standard way of reducing overfitting. However, we can also generate a larger variation in flow fields seen during training by applying a additional spatial transform to the second image, so that we see a larger range of displacements between images during training (the spatial transform can be seen as just another example of ïmage motion). I think this is described first in the original FlowNet paper.

bryanyzhu commented 6 years ago

Thanks a lot, I see it now.