---

name: pysot description: "Operate PySOT visual object tracking workflows: configs, model/tracker APIs, demo/test inference, training data preparation, and toolkit evaluation." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

PySOT repo skill

Use this skill when the task is about PySOT, SiamRPN/SiamRPN++/DaSiamRPN/SiamMask visual object tracking, PySOT experiment YAMLs, PySOT tracker snapshots, training data preparation, or the PySOT evaluation toolkit.

This skill is an operating router. It gives safe preflights and routes to focused sub-skills; it does not claim that full tracking, training, or evaluation can run without the user's snapshots, datasets, video/display setup, and compatible PyTorch/CUDA environment.

First checks

1. Identify the user's task family: config/model, tracking inference, training/data, or evaluation/results.

2. Confirm whether the task is a safe preflight or a full native workflow. Safe preflights may run without model/data assets; full workflows usually need external files and sometimes CUDA.

3. If the environment itself is uncertain, read references/environment-and-install.md and run:

   python scripts/check_env.py --repo-root <pysot-checkout>
   

   Add a config/model construction smoke only when the user has a config:

   python scripts/check_env.py --repo-root <pysot-checkout> --config <config.yaml> --model-smoke
   

4. If the current repository version may differ from this skill, read references/repo-provenance.md before acting.

Route map

Configs, models, trackers, and model zoo

Read sub-skills/configuration-models/SKILL.md when the user asks to:

• edit or validate an experiment YAML;
• choose a SiamRPN/SiamMask/long-term/AlexNet/ResNet/MobileNetV2 config;
• understand cfg, META_ARC, BACKBONE, RPN, MASK, REFINE, TRACK, or ANCHOR keys;
• instantiate ModelBuilder() or build_tracker(model) safely;
• diagnose config merge, unsupported component type, anchor-count, output-size, or snapshot/config mismatch errors.

Tracking inference, demo, and benchmark test commands

Read sub-skills/tracking-inference/SKILL.md when the user asks to:

• run a webcam/video/image-folder demo;
• validate config/snapshot/video/dataset inputs before tracking;
• use tracker.init(img, bbox) and tracker.track(img) programmatically;
• construct a tools/test.py benchmark run and understand result output paths;
• debug OpenCV GUI/video, missing snapshot, CUDA, bbox/mask output, or tracker API failures.

Training data and distributed training setup

Read sub-skills/training-data/SKILL.md when the user asks to:

• prepare VID, YouTube-BB, DET, or COCO training data;
• understand crop image names, train.json, SubDataset, or TrkDataset expectations;
• validate training config/data paths;
• construct distributed tools/train.py commands, pretrained backbone paths, resume settings, logs, or snapshots;
• debug size not match!, missing annotations/images, distributed launch, or training CUDA issues.

Evaluation toolkit, result layouts, metrics, and hp-search

Read sub-skills/evaluation-toolkit/SKILL.md when the user asks to:

• validate tracker result trees before eval.py;
• evaluate OTB, UAV, NFS, LaSOT, VOT short-term, or VOT2018-LT results;
• choose OPE, accuracy/robustness, EAO, or F1 metric flows;
• diagnose tracker_prefix, missing tracker directories, JSON sidecars, region extension, multiprocessing, or result-file naming;
• plan or interpret hp_search.py outputs.

Cross-cutting guidance

• PySOT's setup.py declares the toolkit distribution and Cython region extension. The main pysot package is normally imported from a checkout/PYTHONPATH or equivalent editable-development context.
• If toolkit.utils.region fails to import, build the extension with a compatible Cython version; references/troubleshooting.md gives the recovery path.
• Do not run full training, benchmark testing, hyperparameter search, dataset crop generation, or metric evaluation as default smoke tests. Use the bundled validators first.
• Keep config/model edits and command construction tied to explicit user-provided paths. Do not silently download model snapshots or benchmark/training datasets.
• Treat CUDA as required for unmodified source training/test/hp-search paths that call .cuda() directly. CPU-safe model construction does not prove those paths are runnable.

Root references and helper

• references/environment-and-install.md — install/import model, dependencies, legacy PyTorch/CUDA expectations, and safe versus full workflow boundaries.
• references/troubleshooting.md — cross-cutting import, extension, dependency, backend, and asset failures.
• references/repo-provenance.md — source commit, package metadata, evidence paths, and refresh baseline.
• references/repo-routing-metadata.json — structured router metadata for managed repo-skill import.
• scripts/check_env.py — safe environment/config/model preflight helper.