Example outputs of SiamFC, SiamRPN, and SiamMask.

PySOT is a high-quality, high-performance codebase designed for visual tracking research. It facilitates the rapid implementation and evaluation of novel research ideas. PySOT includes implementations of the following state-of-the-art visual tracking algorithms:

• SiamFC
• SiamRPN
• DaSiamRPN
• SiamRPN++
• SiamMask

These algorithms are supported by powerful backbone network architectures:

• ResNet (18, 34, 50)
• MobileNetV2
• AlexNet

For further details on these models and architectures, see the References section.

PySOT's evaluation toolkit supports the following datasets:

• OTB2015
• VOT16/18/19
• VOT18-LT
• LaSOT
• UAV123

• Simplify inference procedures via CLI.
• Simplify training setup via CLI.
• Automate data downloads (model weights, datasets, videos).
• Expand documentation.

For detailed plans, refer to our Roadmap Document.

git clone https://github.com/MinLee0210/pysot.git
cd pysot
pip install -r requirements.txt

python -m pysot --model_name="<model_name>" --video_name="<video_name>"

Note: Video paths can be local or URL-based (YouTube links preferred).

For comprehensive details on the technologies and methodologies used in PySOT, please consult the following publications:

• Fast Online Object Tracking and Segmentation: A Unifying Approach - IEEE CVPR, 2019
• SiamRPN++: Evolution of Siamese Visual Tracking with Very Deep Networks - IEEE CVPR, 2019
• Distractor-aware Siamese Networks for Visual Object Tracking - ECCV, 2018
• High Performance Visual Tracking with Siamese Region Proposal Network - IEEE CVPR, 2018
• Fully-Convolutional Siamese Networks for Object Tracking - ECCV Workshops, 2016

• Minh, Le Duc
• Fangyi Zhang
• Qiang Wang
• Bo Li
• Zhiyuan Chen
• Jinghao Zhou

PySOT is released under the Apache 2.0 license.