search

visinf / irr

Iterative Residual Refinement for Joint Optical Flow and Occlusion Estimation (CVPR 2019)
Apache License 2.0
194 stars 32 forks source link

About data augmentions and their parameters #10

Closed simon1727 closed 5 years ago

simon1727 commented 5 years ago

Hi!

In the paper, it says that the same geometric and photometric augmentations are implemented as that of Flownet2. However, I noticed that same parameters are distinct comparing to that in https://github.com/lmb-freiburg/flownet2, and an additional Hue transform is implemented, while the 'eigen vector chromatic' transform is not implemented. Are those augmentations and corresponding parameters in this code what were applied to the data during training and finetuning?

Thanks!

hurjunhwa commented 5 years ago

Hi,

Yes, technically I have to admit that it's not completely the same. Instead of reproducing the Caffe source code into pytorch, we used our own version of photometric augmentation, which turned out to make no big difference when comparing the EPE of our baselines (FlowNetS, PWC-Net) in Pytorch with the original baselines in Caffe.

Yes all parameters in the code are exactly the one used during training and fine-tuning.

simon1727 commented 5 years ago

Got it, thanks!