/flowattack

Attacking Optical Flow (ICCV 2019)

Primary LanguagePythonOtherNOASSERTION

Flow Attack

This is an official repository of

Anurag Ranjan, Joel Janai, Andreas Geiger, Michael J. Black. Attacking Optical Flow. ICCV 2019.

[Project Page] [Arxiv]

Known Issues

  • To obtain the batch, use the learning rate of 1e3 and 1e4. For each learning rate, run at least five different trials for 30 epochs.
  • The best patch for FlowNetC was obtained with LR of 1e3 and for FlowNet2 was obtained with LR of 1e4.

Prerequisites

Python3 and pytorch are required. Third party libraries can be installed (in a python3 virtualenv) using:

pip3 install -r requirements.txt

Install custom cuda layers for FlowNet2 using

bash install_flownet2_deps.sh

Preparing training data

Download the KITTI dataset using this script provided on the official website, and then run the following command.

python3 data/prepare_train_data.py /path/to/raw/kitti/dataset/ --dataset-format 'kitti' --dump-root /path/to/resulting/formatted/data/ --width 1280 --height 384 --num-threads 1 --with-gt

For testing optical flow ground truths on KITTI, download KITTI2015 dataset.

Pretrained Models

Download the pretrained models for FlowNetC, FlowNet2, PWC-Net. The pretrained models for SPyNet and Back2Future are provided with this repository.

Generating Adversarial Patches

White-Box Attacks

In the White-Box attacks we optimize a patch for a single network. We use gradient descent as described in the paper. Use the following command to generate an adversarial patch for a specific network architecture using the prepared dataset:

python3 main.py --data [Path to prepared dataset] --kitti-data [Path to KITTI 2015 test set] --flownet [FlowNetS|FlowNetC|FlowNet2|PWCNet|Back2Future|SpyNet] --patch-size 0.10 --name [Name of the experiment]

The patch size is specified in percentage of the training image size (default: 256). All other arguments such as the learning rate, epoch size, etc are set to the values used in our experiments. For details please check main.py

Acknowledgements