This project is based on the Attacking Optical Flow. Authors shows that adversarial attacks can corrupt optical flow estimation and tests different architectures. In general latest approaches have more stable behavior.
Code is based on mmflow framework.
[x] Collect models for optical flow estimation and make it runnable to test
[x] Reproduce adversarial attack as it shown in the paper
[] Attempt to optimize specific adversarial patch for selected models
Sintel dataset were selected for testing and reproducing.
For each model were 3 runs without any patches, with white patch, with universal patch.
EPE was used to measure impact of adding patches.
| Model | Baseline | White Patch | Adversarial patch |
|---|---|---|---|
| FlowNet | 4.5552 | 6.3561 | 6.578 |
| PWCnet | 2.012 | 4.308 | 4.437 |
| RAFT | 1.471 | 3.731 | 3.844 |
Link to gdrive with .mp4
[] Test on more datasets
[] Attempt to optimize specific patch