In the Appendix of your paper, I do not see much difference between the results with or without RefineNet, but RefineNet increase much computational cost.
So the only reason to use RefineNet is to get higher score (like PSNR, SSIM)?
Probably. We found that using only optical flow to warp the image to produce results produced a sharper effect. If we carefully design and optimize the lpips metric, we can get a more practical model.
In the Appendix of your paper, I do not see much difference between the results with or without RefineNet, but RefineNet increase much computational cost. So the only reason to use RefineNet is to get higher score (like PSNR, SSIM)?