ADFactory: An Effective Framework for Generalizing Optical Flow With NeRF

Scale-flow (Ours) Scale-flow MDFlow GMFlow RAFT 𝑥 𝑦 Self-supervised :Trained by our automated data factory Supervised: Trained by existing synthetic and real-world datasets Self-supervised: Trained by photometric loss Figure 1. Zero-Shot Generalization Results in Real World. On top is Scale-flow using our data factory scheme to estimate optical flow results in real-world scenarios. Below is a comparison with existing advanced supervised and self-supervised methods. Our method shows unprecedented accuracy and clarity. Moreover, our fully automated data factory requires no manual intervention and only utilizes photos captured by a monocular camera to train optical flow tasks.