Skip to main content

U-Net's 3D Time Travel: Filling in Occluded Frames

·44 words·1 min · Download pdf

This is impressive performance achieved by rather simple design of U-Net using 3D convolutions. They train this using few starting and ending frames and having network predict the middle frames. The network does unusually good towards dealing with complex scenes with occlusions. https://x.com/tarun_05/status/1610940004263657474

Discussion