DeepLabCut (Claude Skill)

Walks Claude through a full markerless animal pose-estimation project in DeepLabCut, from frame extraction and keypoint labeling to network training, video inference, and behavioral quantification.

   
Type Claude Skill
Supplier Awesome Cognitive and Neuroscience Skills (community OSS, MIT)
Availability GA — one of ~40 research skills in the collection (last upstream push 2026-07-13)
Pricing Free / OSS (MIT)
Capabilities Read/Write — Claude writes and runs DeepLabCut Python locally (Bash); not an MCP tool
Verified works · 2026-08-03
Security cleared · 2026-08-03 — MIT confirmed, repo transferred to NeuroAIHub org (old path still redirects), no advisories

How to install

  • Claude Code — plugin marketplace (installs all skills in the collection): 
    /plugin marketplace add HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills
/plugin install awesome-cognitive-and-neuroscience-skills@awesome-cognitive-and-neuroscience-skills

    Restart Claude Code afterwards. The skills are description-activated — there is no slash command to invoke; ask a pose-estimation question and Claude loads the skill.

  • Claude Code — single-skill alternative, if you do not want the other ~39 skills: 
    git clone https://github.com/HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills
cp -r awesome_cognitive_and_neuroscience_skills/skills/deeplabcut ~/.claude/skills/

    (Project-scoped alternative: copy into .claude/skills/ instead of ~/.claude/skills/.)

  • Underlying software — the skill drives DeepLabCut in your own Python environment, so install it first. PyTorch must be installed separately and matched to your CUDA version (see pytorch.org): 
    pip install "deeplabcut[gui]"

    The [gui] extra is what provides the keypoint-labeling interface; omit it for headless training/inference only.

What it does

Covers the nine-stage DeepLabCut pipeline:

  • Project setup — create the project structure and register video paths.
  • Frame extraction — sample training frames by k-means clustering so labeled frames span the behavioral range.
  • Annotation — label keypoints in the GUI, then convert annotations into a training dataset.
  • Training — train ResNet-50/101, MobileNet v2, or EfficientNet-b0 backbones (default maxiters 100,000, saveiters 5,000, batch size 8 — drop to 4 or 2 when GPU memory is tight).
  • SuperAnimal pretrained models — zero-shot inference without labeling, with a recommended starting scale list of [200, 300, 400].
  • Evaluation — test-set error before committing to full-video inference.
  • Inference, visualization, export — predict poses across videos, overlay tracked keypoints, and export to CSV.
  • Multi-animal tracking — identity assignment on top of a FasterRCNN (fasterrcnn_resnet50_fpn_v2) detector; 3D pose estimation from calibrated multi-camera setups.

Primary use cases: rodent and non-human-primate behavioral tracking, gait and kinematics quantification, multi-animal social-behavior scoring.

Notes

AI-generated content — verify before use. The collection’s README states that all skills “were generated by AI from academic papers and other publicly available sources” and that “the content has not been individually verified by human domain experts.” This skill’s front-matter carries review_status: ai-generated and lists Claude (AI-assisted) as author. Treat the parameter defaults above as starting points and check them against the DeepLabCut documentation before running an experiment.

DeepLabCut itself is separately licensed (LGPL-3.0) and is not redistributed by the skill — the MIT grant covers the skill text only. GPU training is effectively required: CPU-only training of 100,000 iterations is impractical. H.264 .mp4 or AVI are the recommended input container formats.

Sources

Installed this tool?

Share feedback — install path, OS, errors, workarounds. The form opens with this tool pre-selected and a link back to this page.