Open-DDVM: A Replica and Extension of Diffusion Mannequin for Optical Stream Estimation
Authors: Qiaole Dong, Bo Zhao, Yanwei Fu
Summary: Not too long ago, Google proposes DDVM which for the primary time demonstrates {that a} normal diffusion mannequin for image-to-image translation process works impressively nicely on optical movement estimation process with none particular designs like RAFT. Nevertheless, DDVM remains to be a closed-source mannequin with the costly and personal Palette-style pretraining. On this technical report, we current the primary open-source DDVM by reproducing it. We examine a number of design selections and discover these vital ones. By coaching on 40k public information with 4 GPUs, our copy achieves comparable efficiency to the closed-source DDVM. The code and mannequin have been launched in https://github.com/DQiaole/FlowDiffusion_pytorch
