original paper: https://arxiv.org/pdf/1910.10053.pdf

All models were trained on FlyingChairs dataset. A set of 1000 test images was also selected from FlyingChairs. Patch size is 50x50 pixels.

• select and download models with checkpoints (flownetc, pwc, raft)
• how to download and add adversarial patch
• create repo
• test EPE from mmflow (https://github.com/open-mmlab/mmflow/blob/master/mmflow/core/evaluation/metrics.py)
• add how to run
• visualize diffreence between predicted uattacked and patched input
• metrics all, without patch, inside patch only
• average runs with different seeds
• add automatic result table generation
• check if model weights are identical
• universal patch was not optimised for raft