SAOT

The official implementation of the paper Saliency-Associated Object Tracking (accepted by ICCV2021).

Updates

Support PyTorch1.7 and CUDA11.0.

Support DDP.

Update the training settings for better performance.

Training and testing our model with the new settings using a single GPU, the performance on LaSOT reaches 0.631 (AUC).

Performance reported in the paper

Dataset	LaSOT (AUC)	GOT-10K (AO)	OTB2015 (AUC)	VOT2018 (EAO)
SAOT	0.616	0.640	0.714	0.501
Ocean	0.560	0.611	0.684	0.489
DiMP	0.568	0.611	0.686	0.440

Real-time Speed

SAOT runs at about 29 FPS on an RTX 2080 GPU with PyTorch 1.1 and CUDA 10.

Model Zoo & Raw Results

The pre-trained model and raw results are provided in model zoo and raw results, respectively.

Installation

Clone the GIT repository.

git clone https://github.com/ZikunZhou/SAOT.git

Clone the submodules.

In the repository directory, run the commands:

git submodule update --init  

Install dependencies

Please note that using a new version pytorch may cause unexpected issues. Following the detailed installation instructions to install the dependencies.

Testing

Download the pre-trained networks. Activate the conda environment and run the script pytracking/test_saot_fs.py.

conda activate SAOT
cd pytracking
python test_saot_fs.py saot saot_otb --dataset OTB2015

Training

Download the training datasets. Activate the conda environment and run the script train.py.

conda activate SAOT
python train.py --train_module dimp --train_name saot --mode multiple --nproc_per_node 1

Citation

Please cite the following publication, if you find the code helpful in your research.

@InProceedings{Zhou_2021_ICCV,
    author    = {Zhou, Zikun and Pei, Wenjie and Li, Xin and Wang, Hongpeng and Zheng, Feng and He, Zhenyu},
    title     = {Saliency-Associated Object Tracking},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {9866-9875}
}

Acknowledgments

Thanks for the PyTracking and Pysot libraries, which helps us to quickly implement our ideas.

Thanks for Kaige Mao for the helps about the updates.

Contact

Please feel free to contact me (Email: [email protected]).