Hello, authors, I really like your work.
There are two details in the code that confuse me.

1. What does pe_deformed_grid = (deformed_grid + 0.3) / 1.6 do in train.py? How did we get this coefficient.
2. In video_dataset.py, flow_tensor[..., 0] /= W_, flow_tensor[..., 1] /= H_ looks like it's inverted. Isn't flow_tensor[..., 0] horizontal? Also, aren't the optical flow outputs already normalized? What does this step do?