ml-lab / splt

`Skimming-Perusal' Tracking: A Framework for Real-Time and Robust Long-term Tracking

`Skimming-Perusal' Tracking: A Framework for Real-Time and Robust Long-term Tracking

This is the official code for the ICCV 2019 paper[Google][Baidu(oj49)]. This code has been tested on

• RTX 2080Ti
• CUDA 10.0 + cuDNN 7.6 / CUDA 9.0 + cuDNN 7.1.2
• Python 2.7
• Ubuntu 18.04.2 LTS

Please cite our paper if you find it useful for your research.

@inproceedings{ iccv19_SPLT,
    title={`Skimming-Perusal' Tracking: A Framework for Real-Time and Robust Long-term Tracking},
    author={Yan, Bin and Zhao, Haojie and Wang, Dong and Lu, Huchuan and Yang, Xiaoyun},
    booktitle={IEEE International Conference on Computer Vision (ICCV)},
    year={2019}
}

Installation

• Create anaconda environment:

conda create -n SPLT python=2.7
conda activate SPLT

• Clone the repo and install requirements:

git clone https://github.com/iiau-tracker/SPLT.git
cd <path/to/SPLT>
pip install -r requirements.txt

• CUDA and cuDNN:

conda install cudatoolkit=10.0
conda install cudnn=7.6.0

# or CUDA 9.0 + cuDNN 7.1.2 for TensorFlow  < 1.13.0
conda install cudatoolkit=9.0
conda install cudnn=7.1.2

Models

Model	Size	Google Drive	Baidu
SiamRPN	215 MB	model.ckpt-470277	Mirror
Verifier	178 MB	V_resnet50_VID_N-65624	Mirror
Skimming	24 MB	Skim	Mirror

• extract model.ckpt-470277 to ./RPN
• extract V_resnet50_VID_N-65624 to ./Verifier
• extract Skim to ./Skim

Demo

# modify 'PROJECT_PATH' in 'demo.py' 
python demo.py

Evaluation on VOT

start from RPN_Verifier_Skim_top3.py

• modify PROJECT_PATH in RPN_Verifier_Skim_top3.py
• add set_global_variable('python', 'env -i <path/to/anaconda/envs/SPLT/bin/python>'); to configuration.m

raw resluts (vot-toolkt version 6.0.3)

Train

Train the Verifier(optional)

Download ResNet50 model pretrained on IMAGENET.Then put extracted ckpt file in train_Verifier/lib

cd train_Verifier/experiments
# modify paths in classify.py
python classify.py
# modify paths in triplet_pairs.py
python triplet_pairs.py
# modify paths in train_multi_gpu.py
python train_multi_gpu.py

Train the Skimming(optional)

cd train_Skim
# modify paths in classify.py
python classify.py
# modify paths in skim_data.py
python skim_data.py
# modify paths in train_skim.py
python train_skim.py