I'm not familiar with Optical Flow research, and I'm sure you already know, but are you familiar with the U^2-Net architecture (a new variation of U-Nets)? I o

I'm not familiar with Optical Flow research, and I'm sure you already know, but are you familiar with the U^2-Net architecture (a new variation of U-Nets)? I only bother mentioning it at the risk of being redundant because it's new and I know how easy it is to miss stuff in the fire hose that is machine learning research papers. Its results for their specific task, salience mapping, are incredible for such a simple and lightweight architecture. And from a skim of Optical Flow and your example it seems like it might apply there as well. It should be able to grab the long-ranging features necessary for Optical Flow. Do you have a public dataset that you can share or only a private one? Kinda looks like you built a dataset using Skyrim which ... is brilliant.