Cooling-Shrinking Attack (CSA)

The official implementation for CVPR2020 Paper Cooling-Shrinking Attack: Blinding the tracker with imperceptible noises

Demos

Demos for Cooling-Shrinking Attack.

Please cite our work as follows, if you find it helpful to your research. :)

@inproceedings{CSA-CVPR2020,
author = {Bin Yan and Dong Wang and Huchuan Lu and Xiaoyun Yang},
title = {{Cooling-Shrinking Attack: Blinding} the Tracker with Imperceptible Noises},
booktitle = {CVPR},
year = {2020}
}

Installation

This code has been tested on the following environment:
       NVIDIA RTX-2080Ti
       Ubuntu 16.04
       CUDA 10.0

Clone the repository

git clone https://github.com/MasterBin-IIAU/CSA.git
cd <Project_name>

Create Environment

conda create -n CSA python=3.6
source activate CSA
conda install pytorch=1.0.0 torchvision cuda100 -c pytorch
pip install -r requirements.txt
conda install pillow=6.1

Prepare the training set (optional)

1. Download the training set of GOT-10K.
2. Then change 'got10k_path' and 'save_path' in Unified_GOT10K_process.py to yours.
3. Finally, run the following script.
   (it takes a long time. After running it, you can do the next steps :)

python Unified_GOT10K_process.py

Download pretrained models

1. SiamRPN++(Model_Zoo)
   Download siamrpn_r50_l234_dwxcorr and siamrpn_r50_l234_dwxcorr_otb
   Put them under pysot/experiments/<MODEL_NAME>
2. Perturbation Generators
   Download checkpoints you need, then put them under checkpoints/<MODEL_NAME>/
   (Google Drive, Baidu[Extraction code: 98rb])

Set some paths

Step1: Add pix2pix and pysot to environment variables

sudo gedit ~/.bashrc
# add the following two lines to the end
export PYTHONPATH=<CSA_PATH>:$PYTHONPATH
export PYTHONPATH=<CSA_PATH>/pysot:$PYTHONPATH
export PYTHONPATH=<CSA_PATH>/pix2pix:$PYTHONPATH
# close the file
source ~/.bashrc

step2: Set another paths

1. Gather testing datasets
   create a folder outside the project folder as <DATASET_ROOT>
   then put soft links for OTB100, VOT2018 and LaSOT into it
2. Set 'project_path_' and 'dataset_root_' Open common_path.py, go to the end
   project_path_ = <CSA_PATH>
   dataset_root_ = <DATASET_ROOT> train_set_path_ = <TRAIN_SET_PATH>

Training (Optional)

Option1: Use Default Settings
Train a generator for attacking search regions (Cooling+Shrinking)

python train1.py # See visualization in http://localhost:8097/

Train a generator for attacking search regions (Only Cooling)

python train0.py # See visualization in http://localhost:8096/

Option2: Change Settings
       If you want to train other models (like the generator for attacking the template), you can change the lines 23 and 24 in pix2pix/options/base_option0.py (or base_option1.py). In specific, modify the default values to 'G_template_L2_500' (or 'G_template_L2_500_regress'). Then run python train0.py or python train1.py
Option3: Train Your Own Models
Step1: Create a new python file under pix2pix/models.
       You can copy a file that belongs to this folder, then develop based on it. Note that the class name must match the filename.
Step2: Change default values and train (Do as instructions in Option2)

Testing

open common_path.py, choose the dataset and siamese model to use.
open GAN_utils_xx.py, choose the generator model to use.
cd pysot/tools
run experiments about attcking search regions

python run_search_adv0.py # or run_search_adv1.py

run experiments about attacking the template

python run_template_adv0.py # or run_template_adv1.py

run experiments about attacking both search regions and the template

python run_template_search_adv0.py # or run_template_search_adv1.py