I use the FlowNet-S model described in the paper: "FlowNet: Learning Optical Flow with Convolutional Networks" to generate our Optical Flow images.
I have converted the weights from the NVIDIA FlowNet2 models from PyTorch to Gluon.
Please download the params from my Google Drive,
and place it in this directory (you will only need FlowNet2-S_checkpoint.params).
To generate the flow images just execute the run.py script. I will also allow for the download of the flow images
directly soon.