take the example of multiplying the last layer by 5.0 as an example, the output of the flow is 1/4 in height and width, and in training we divide the flow by 20. Therefore, we need to multiply by 20/4 = 5.
you may have a try by fine tuning pwcnet on your own HD dataset, and take a look at the output.
The HD1K data are of higher resolution (2560x1080) than the training data and the trained model works on the HD1K data. The only constraint is that you need a GPU with large memory (>= 16G), such as Volta 100, to run on full HD size.
Hi, Nice work and your code works well Thanks!
Just a simple question about training and testing
Is there any reason why you multiply (0.65 1.25 2.5 5.0) to flow not (1.0 2.0 4.0 8.0). ( maybe this parameter come from flownet implementaion?)
Can model predict flow larger than trained resolution images? ( eg Full HD size )
waiting your reply Thanks!