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

Pass in two identical images #21

Open cty-xiaobai opened 3 years ago

cty-xiaobai commented 3 years ago

Passing two identical images to the network, why the pixel displacement is not 0

gengshan-y commented 3 years ago

Does it work properly on the example images?

cty-xiaobai commented 3 years ago

Does it work properly on the example images?

It can run normally on the example image, but if two identical images are passed in, the pixel displacement is not 0

gengshan-y commented 3 years ago

Can you give more details about the error? Are the displacements close to zero (<1px), or very large? Is this specific to some data, or it exist for many pairs.

If the error is small, I could image this happening due to dataset bias, considering VCN is still a data-driven method and is not trained on pairs with zero motion. If the error is not negligible, it would be interesting to look into. A way to address this is to add zero-motion pairs to the training data.