Real Time Visual Tracking using Spatial-Aware Temporal Aggregation Network

By Tao Hu, Lichao Huang, Han Shen.

The code for the official implementation of paper SATA.

The code will be made publicly available shortly.

Motivation
Performances
Environment Setup
Training & Inference
Acknowledgement
Bibtex

Motivation

Performances

Results

SATA obtains a significant improvements by feature aggregation

The OPE/TRE/SRE results on OTB GoogleDrive.

OTB2013:

OTB2015:

Note:

The results are better than reported in the paper because we modify some hyper-parameters.
The results could be slightly different depending on the running environment.

Environment Setup

Env Requirements:

Linux.
Python 3.5+.
PyTorch 1.3 or higher.
CUDA 9.0 or higher.
NCCL 2 or higher if you want to use distributed training.

Download ILSVRC2015 VID dataset from ImageNet, extract files and put it in the directory you defined.
After the download, link your datasets to the current directory, like,
```
cd data
ln -s  your/path/to/data/ILSVRC2015 ./ILSVRC2015
```

Split training and validation set. For example, I use MOT17-09 as validation set, and others video as training set.

./ILSVRC2015
├── Annotations
│   └── VID├── a -> ./ILSVRC2015_VID_train_0000
│          ├── b -> ./ILSVRC2015_VID_train_0001
│          ├── c -> ./ILSVRC2015_VID_train_0002
│          ├── d -> ./ILSVRC2015_VID_train_0003
│          ├── e -> ./val
│          ├── ILSVRC2015_VID_train_0000
│          ├── ILSVRC2015_VID_train_0001
│          ├── ILSVRC2015_VID_train_0002
│          ├── ILSVRC2015_VID_train_0003
│          └── val
├── Data
│   └── VID...........same as Annotations
└── ImageSets
    └── VID

Download pre-train model for backbone, and put it in ./pretrainmodel

cd ./pretrainmodel
wget  https://download.pytorch.org/models/resnet50-19c8e357.pth

Install others dependencies. Our environment is PyTorch 1.3.0+cu92, torchvision 0.4.1+cu92.
```
pip install -r requirements.txt
```

Training & Inference

train model:

#  Multi GPUs (e.g. 4 GPUs)
cd ./experiments/train_tsr
CUDA_VISIBLE_DEVICES=0,1,2,3 python ../../scripts/train.py --cfg config.yaml

inference for online version:

cd ./experiments/test_tsr
python ../../scripts/test.py --cfg config.yaml

Acknowledgement

The work was mainly done during an internship at Horizon Robotics.

Bibtex

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follows.

@article{hu2019real,
  title={Real Time Visual Tracking using Spatial-Aware Temporal Aggregation Network},
  author={Hu, Tao and Huang, Lichao and Liu, Xianming and Shen, Han},
  journal={arXiv preprint arXiv:1908.00692},
  year={2019}
}

License

For academic use, this project is licensed under the 2-clause BSD License - see the LICENSE file for details. For commercial use, please contact the authors.

ecart18/SATA