One PyTorch shop for the Siamese tracking dynasty

What it does

PySOT is a Python research platform built by SenseTime for single-object visual tracking. It packages implementations of several major Siamese-network trackers—including SiamFC, SiamRPN, SiamRPN++, DaSiamRPN, and SiamMask—on top of backbones like ResNet, MobileNetV2, and AlexNet. The repository also includes a Python port of the standard evaluation toolkit, supporting benchmarks such as OTB2015, VOT16/18/19, VOT18-LT, LaSOT, and UAV123.

The interesting bit

The project treats flexibility as a first-class feature: backbones are swappable, and the evaluation pipeline is baked in, so the gap between “implement a new idea” and “compare it on five datasets” is mostly just code. That is the unglamorous part most research codebases skip.

Key highlights

• Ships with five Siamese tracker implementations and a model zoo of pretrained baselines.
• Supports six standard tracking benchmarks through its bundled evaluation toolkit.
• Backbones (ResNet, MobileNetV2, AlexNet) are modular and designed to be extended.
• Released under Apache 2.0 by the SenseTime Video Intelligence Research team.

Caveats

• The README notes that PYTHONPATH must be set manually and a C extension (region) must be built, suggesting the packaging has rough edges.
• It is explicitly aimed at research; production deployment characteristics are not discussed.

Verdict

Worth a look if you are building or benchmarking Siamese trackers and want a shared PyTorch foundation to stand on. Skip it if you need a polished, pip-installable application rather than a research sandbox.