This is the official implementation of the SiamTPN (WACV2022). The tracker intergrates pyramid feature network and transformer into Siamese network, achieving state-of-the-art performance (better than DiMP) while runing 30 FPS on a single CPU. The tracker optimized with ONXX and openvino could run at 45 FPS on cpu end, leading promising performance when deploying on drones for tracking.
[Paper] [Raw Results] [Drone Tracking Videos] [Models]
change the path in lib/train/admin/local.py to your data location
# Distributed training withh 4 nodes
python -m torch.distributed.launch --nproc_per_node 4 tools/run_training.py --config shufflenet_l345_192
# single gpu training for test purpose
python tools/run_training.py --config shufflenet_l345_192
change the path in lib/test/evaluation/local.py to your data location
# Download the pretrain model and put it under ./results/checkpoints/train/SiamTPN/ folder
python tools/test.py siamtpn shufflenet_l345_192 --dataset_name got10k_val --debug 1 --cpu 1 --epoch 100 --sequence GOT-10k_Val_000001
The debug mode will show tracking results, more details refer to tools/test.py
Currently, onnx only support cpu version
First, you need to install onxx and onxxruningtime:
pip install onxx
# for onxx runining time, download the openvino version from release [page](https://github.com/intel/onnxruntime/releases/tag/v3.1) and install with
pip install onnxruntime_openvino-1.9.0-cp37-cp37m-linux_x86_64.whl
# please refer the [page](https://github.com/intel/onnxruntime/releases/tag/v3.1) for openvino installation details.
# Download the converted onnx model and put it under ./results/onnx/ folder
# or conver your own model with
python tools/onnx_search.py
python tools/onnx_template.py
python tools/test.py siamtpn_onnx shufflenet_l345_192 --dataset_name got10k_val --debug 1 --cpu 1 --epoch 100 --sequence GOT-10k_Val_000001
If you find this repo useful, please cite with
@article{xing2021siamese,
title={Siamese Transformer Pyramid Networks for Real-Time UAV Tracking},
author={Xing, Daitao and Evangeliou, Nikolaos and Tsoukalas, Athanasios and Tzes, Anthony},
journal={arXiv preprint arXiv:2110.08822},
year={2021}
}
Our code is implemented based on the following libraries: