Issue with Flow representations

[saliknadeem]

Hi,

I noticed an issue with the way you have handled the Flow calculations in the dataloader. In the paper they mentioned that they downsampled the flow array using an 8*8 kernel using average pooling. You have taken a center crop to downsample, which is not going to give the right results.

Secondly, the paper mentioned to multiply the resulting flows by 50. I'm not sure what that is meant to do. Multiplying the flow array with 50 just increases the values by a lot. But if they meant upsample it by 50 then its like 28*50=1400px which becomes huge. So I'm not sure what that part does.

[minostauros]

Hi @saliknadeem,

Thank you very much for your analysis.
At this moment I could not validate those issues myself, so please create pull requests if you'd like to join, or you could just leave a link to your commit of your own repo.

For the second issue, I also wondered the effectiveness of the multiplication, but it might be just to scale up the range of flow values to be used with RGB values (just my guess).