siamattack's Introduction

环境

pytorch1.4_python3.7

训练

python main/train.py --config experiments/siamfcpp/train/got10k/siamfcpp_googlenet-trn.yaml -r /home/etvuz/projects/adversarial_attack/video_analyst/models/siamfcpp/siamfcpp-googlenet-got-md5_e182dc4c3823427022eccf7313d740a7.pkl

训练所有数据

source activate pytorch1.4_python3.7
export CUDA_VISIBLE_DEVICES=1,2,3
cd /home/etvuz/projects/adversarial_attack/video_analyst
python main/train.py --config experiments/siamfcpp/train/fulldata/siamfcpp_googlenet-trn-fulldata.yaml -r /home/etvuz/projects/adversarial_attack/video_analyst/models/siamfcpp/siamfcpp-googlenet-got-md5_e182dc4c3823427022eccf7313d740a7.pkl

可视化

将 txt 跟踪结果可视化

main/visualize_txt_result.py

训练过程可视化

videoanalyst.utils.visualize_training

搜索图像可视化

videoanalyst.utils.visualize_search_img

生成相对于原始图像的 Fake Ground Truth

python videoanalyst/pipeline/utils/generate_patch_anno.py
数据位于 /home/etvuz/projects/adversarial_attack/patch_anno

测试

生成 Fake Ground Truth

[修改相应路径] python videoanalyst/pipeline/utils/generate_patch_anno.py
数据位于 /home/etvuz/projects/adversarial_attack/patch_anno/[dataset_name]

使得测试代码正确读入 FGT

got_benchmark_helper.py/PipelineTracker/track():dataset_name = [dataset_name]

设置是否进行攻击

siamfcpp_track.py/self.do_attack = ...

设置 iteration num

siamfcpp_track.py/self.loop_num = ...

在配置文件中指定正确的预训练模型

/home/etvuz/projects/adversarial_attack/video_analyst/models/siamfcpp/siamfcpp-googlenet-got-md5_e182dc4c3823427022eccf7313d740a7.pkl

进入 conda 环境

source activate pytorch1.4_python3.7
cd /home/etvuz/projects/adversarial_attack/video_analyst

测试

python main/test.py --config experiments/siamfcpp/train/got10k/siamfcpp_googlenet-trn.yaml
python main/test.py --config experiments/siamfcpp/test/otb/siamfcpp_googlenet-otb.yaml
python main/test.py --config experiments/siamfcpp/test/lasot/siamfcpp_googlenet-lasot.yaml

评估

[修改 dataset_name/loop_num] python main/eval.py

实时评估

python main/realtime_eval.py

siamattack's People

Contributors

Stargazers

Watchers

siamattack's Issues

关于训练问题

您好，我想问下我直接运行readme中的训练命令：

python main/train.py --config experiments/siamfcpp/train/got10k/siamfcpp_googlenet-trn.yaml -r /home/etvuz/projects/adversarial_attack/video_analyst/models/siamfcpp/siamfcpp-googlenet-got-md5_e182dc4c3823427022eccf7313d740a7.pkl

epoch直接就是20了，epoch20中生成的patch便可以实现很好的攻击效果，不需要更多的epoch。请问我这样做正确嘛？没有用到您训练好的攻击模型吧？siamfcpp-googlenet-got-md5_e182dc4c3823427022eccf7313d740a7.pkl应该是tracker的训练好的模型吧。

非常感谢您的解答！