wattanapong / DFA

Reference

The DFA was published in Neurocomputing 2022. If our work is benefit to you please cite this reference.

Diminishing-feature attack: The adversarial infiltration on visual tracking. Suttapak, Wattanapong, Jianfu Zhang, and Liqing Zhang. Neurocomputing 509 (2022): 21-33. Neurocomputing.

Resource requirements

We test with Nvidia RTX 2070s. This project was implemented under Pytorch 1.4.0 and CUDA 11.2.

PySOT

This project was forked from pysot repository. If you need to reproduce training our model, you need to setup pysot repository first.

Installation

create new environment from requirements.yaml

conda env create --file requirements.yml
conda activate dfa

python setup.py build_ext --inplace

Download SiamRPN++ pretrained

• We also suggest to download this pretrained from original repository.
• Change pretrained name as siamrpn_r50_otb_model.pth

Download Our pretrained DFA model

• We only support google site, access here.

Download dataset

• We only share biker frames and json of OTB100 dataset . You can download full download from providing dataset source.

Testing

Beforehand testing this step, you need to setup the pretrained path, dataset path and save path.

cd DFA/experiments/siamrpn_r50_l234_dwxcorr_otb
bash test_dfa.sh

Example results of our DFA model

You can access: 1. DFA + SiamRPN++ 2. DFA transferability

License

PySOT is released under the Apache 2.0 license.