---

name: variant-catalog description: "Route natural tracker-variant requests in SiamTrackers, distinguish maintained NanoTrack from alternative snapshots and references, and select a safe evidence-bounded workflow." disable-model-invocation: true metadata: disco-role: operating license: NO_LICENSE

SiamTrackers variant catalog

Use this skill when a request names an alternative tracker, asks which tracker to choose, or mixes tracker selection with build, data, training, test, evaluation, or deployment requirements. The collection's maintained end-to-end workflow is NanoTrack. The other entries are useful historical or comparative snapshots, but their presence does not prove that their dependencies, checkpoints, data, or full workflows are still runnable.

Route first

1. Extract the request's objective: smallest/fastest deployment, baseline comparison, anchor-free box tracking, RPN tracking, segmentation, transformer tracking, model updating, training, benchmark evaluation, or export.
2. Ask whether the user has a current source checkout, authorized data and model artifacts, and a compatible isolated environment. If any is missing, provide a plan and an evidence boundary rather than inventing a runnable command.
3. Choose NanoTrack for the default maintained workflow, especially for lightweight CPU/mobile/embedded inference, training on the collection's documented cropped tracking data, or the collection's documented ONNX/NCNN direction. Route detailed NanoTrack implementation to the inference, training, evaluation, or export sibling skill as appropriate.
4. Choose an alternative only when the request names it or its distinctive research behavior matters. Use the comparison in variant-overview.md, then apply the build and legacy-runtime gates in legacy-workflows.md.
5. If the request is only “what is included?”, answer from the catalog and label every entry by evidence level. Do not turn a README link, result archive, or historical result table into an availability claim.

Quick selection guide

Need	First route	Important caveat
Maintained lightweight tracker	NanoTrack	Detailed model/API/export work belongs to sibling skills.
Classic fully-convolutional baseline	SiamFC	Snapshot has train/test/eval code, but no uniform setup contract.
RPN baseline or deeper RPN	SiamRPN, then SiamRPNpp	Pysot-like snapshots need their own config/checkpoint pairing.
Anchor-free box tracker	SiamBAN or SiamCAR	Source snapshots share legacy Cython/toolkit conventions.
Deeper/wider FC or RPN	SiamDW-FC or SiamDW-RPN	Older training/data layout and no shared region setup file.
Segmentation plus tracking	SiamMask	Mask datasets and mask-specific model/config are mandatory.
Learned template update	UpdateNet	It is an updater layered on DaSiamRPN, not a standalone replacement.
Transformer tracker	TrTr	GPU and transformer-compatible legacy stack are expected.
SiamFC++ framework variants	SiamFCpp-pysot or video_analyst	These are separate snapshots with different source layouts.
LightTrack/Ocean	Reference only	This checkout does not contain their implementation.

Evidence levels and non-goals

• Maintained complete workflow: NanoTrack has source, configs, launcher families, toolkit, build metadata, and documented train/test/eval/export-shaped paths. Its full run is still gated on user-supplied data and models.
• Substantive snapshot: DaSiamRPN, SiamBAN, SiamCAR, SiamDW-FC, SiamDW-RPN, SiamFC, both SiamFCpp snapshots, SiamMask, both SiamRPN snapshots, SiamRPNpp, TrTr, and UpdateNet contain implementation roots and representative launchers. “Substantive” means source evidence exists, not that a modern environment has passed end-to-end execution.
• Reference-only: LightTrack and Ocean contain a thin README/reference pointer in this collection, not a local implementation or reproducible package. Do not route a build or evaluation request to them as if source were bundled.
• Excluded: SiamFace is a separate face-classification Siamese demo. It is not a single-object visual tracker and is outside this catalog.
• This skill does not perform training, benchmark evaluation, checkpoint recovery, model download, or export. No full tracking/training/evaluation run is claimed by this catalog.

Common gates before any alternative run

• Work from a user-provided, authorized active checkout or package source; do not rely on this skill's original checkout, external official repositories, or archive links. The collection's historical download references are provenance only, not bundled runtime sources.
• Confirm the selected snapshot's source root, config root, model format, expected dataset layout, result directory, and launcher family as one set. Do not mix a config, checkpoint, or toolkit from a different variant merely because names look similar.
• Use an isolated environment. The historical requirements commonly target Python 3.7-era and PyTorch 1.2–1.4-era stacks; treat those pins as compatibility evidence, not installation commands. Prefer a deliberately tested compatibility matrix over blind downgrades.
• Treat a Cython region extension as a build prerequisite when the selected evaluation toolkit imports it. A prebuilt extension compiled for another Python ABI is not proof of importability or evaluation correctness.
• Require actual dataset files, annotations/JSON, and matching checkpoints before test/eval. Empty directory skeletons, result ZIPs, README metrics, or a model filename in a default argument are not sufficient evidence.
• For CUDA or distributed training, check the selected environment, GPU visibility, torch/CUDA compatibility, process launcher, and batch/memory plan before starting. A preparation smoke test is not a tracker test.

Route boundaries

• NanoTrack deep implementation: use inference for stateful model/API behavior, training for configs and training, evaluation for benchmarks, and export for ONNX/mobile/deployment. This catalog only chooses the route.
• Alternative implementation details: stay here for selection, evidence, prerequisites, and legacy troubleshooting; do not reproduce large launcher or model scripts.
• Metrics: historical tables are comparative context only. Report the dataset, protocol, checkpoint, config, and measurement source before treating a number as a result.
• External acquisition: request an authorized local artifact or an approved acquisition procedure from the user. Never claim that a named official repo, BaiduYun archive, password, or thin README is available to a future runtime.

Bundled references

• Variant overview: source-root inventory, evidence labels, comparison matrix, and per-variant routing.
• Legacy workflows: build, data/model, train/test/eval, and deployment patterns distilled from representative bins.
• Troubleshooting: compatibility conflicts, extension failures, artifact mismatches, and safe stop conditions.