search

gengshan-y / VCN

Volumetric Correspondence Networks for Optical Flow, NeurIPS 2019.
https://gengshan-y.github.io/VCN/
MIT License
154 stars 25 forks source link

PWCNet comparison experiment #8

Closed Prinsphield closed 4 years ago

Prinsphield commented 4 years ago

Hi Gengshan,

If I change the model to PWCNet as you provided in the repo and run the code under the same training scheme, can I get the same result in the this paper or the original paper in PWCNet?

Thanks~

gengshan-y commented 4 years ago

You also have to change some hyperparameters according to our supplement Sec 6.1. For example, change learning rate to 1e-4, otherwise the weights will explode.

After changing the hyperparameters, I was able to reproduce the result of PWCNet trained on Chairs and things.