---

name: deeplabcut description: "Toolbox for markerless animal pose estimation with DeepLabCut. Covers single/multi-animal tracking, SuperAnimal pretrained models, 2D/3D pose estimation, keypoint labeling GUI, model training/evaluation, video analysis, and behavioral quantification. Use when the user needs animal pose estimation, behavior tracking, keypoint detection in videos, or mentions DeepLabCut/DLC/SuperAnimal." version: "1.0.0" authors:

• "Claude (AI-assisted)" review_status: "ai-generated" source: "https://github.com/DeepLabCut/DeepLabCut"

---

DeepLabCut — Markerless Animal Pose Estimation

Purpose

DeepLabCut is a Python toolbox for state-of-the-art markerless pose estimation of animals. It uses deep learning to track body parts from videos without physical markers. The library is animal-agnostic, supports both single and multi-animal scenarios, and includes the SuperAnimal family of pretrained models for out-of-the-box inference.

When to Use This Skill

Activate when the user:

• Wants to track animal body parts from video
• Asks about pose estimation, keypoint detection, or behavioral tracking
• Mentions DeepLabCut, DLC, SuperAnimal, or markerless tracking
• Needs to analyze animal movement/kinematics
• Asks about multi-animal tracking or 3D pose reconstruction
• Wants to use pretrained animal pose models

Quick Decision Tree

What does the user need?
├── No labeled data, just want to track animals → SuperAnimal (video_inference_superanimal)
├── Single animal, have labeled data → Standard single-animal pipeline
├── Multiple animals interacting → maDLC (multi-animal pipeline)
├── 3D pose reconstruction → 3D pipeline (calibrate_cameras + triangulate)
└── Just post-process results → filterpredictions, analyzeskeleton

Reference Files (Progressive Disclosure)

Topic	File	When to Read
Standard Pipeline	references/standard-pipeline.md	Full workflow: create project → train → analyze
SuperAnimal & ModelZoo	references/modelzoo.md	Pretrained models, zero-shot inference
Multi-Animal (maDLC)	references/maDLC.md	Tracking multiple interacting animals
3D Pose Estimation	references/3d-pose.md	Triangulation from multiple camera views
Video & Data Utilities	references/utilities.md	Video cropping, format conversion, data export

Installation

# Minimal (headless, no GUI):
pip install deeplabcut

# With GUI (label_frames, refine_labels, SkeletonBuilder):
pip install "deeplabcut[gui]"

# PyTorch must be installed separately:
pip install torch torchvision
# Or for GPU (check pytorch.org for your CUDA version):
conda install pytorch cudatoolkit=11.3 -c pytorch

Verify: python -c "import deeplabcut; print(deeplabcut.__version__)"

Standard Pipeline Overview

import deeplabcut as dlc

# 1. Create project
config_path = dlc.create_new_project(
    "ProjectName", "ExperimenterName", ["/path/to/video.mp4"],
    working_directory="/path/to/projects"
)

# 2. Extract frames for labeling
dlc.extract_frames(config_path, mode="automatic", algo="kmeans", crop=True)

# 3. USER labels frames manually in GUI
# dlc.label_frames(config_path)  # launches the labeling GUI

# 4. Create training dataset from labeled frames
dlc.create_training_dataset(config_path, net_type="resnet_50")

# 5. Train the network
dlc.train_network(config_path, maxiters=100000, saveiters=5000)

# 6. Evaluate
dlc.evaluate_network(config_path, plotting=True)

# 7. Analyze videos (predict poses)
dlc.analyze_videos(config_path, ["/path/to/video.mp4"], videotype=".mp4")

# 8. Create labeled videos (overlay predictions)
dlc.create_labeled_video(config_path, ["/path/to/video.mp4"])

# 9. Export results to CSV
dlc.analyze_videos_converth5_to_csv("/path/to/videoDLC_resnet50_ProjectNameJul9")

Key API Reference

Project Management

Function	Description
create_new_project(project, experimenter, videos, working_directory)	Start a new single-animal project
create_new_project_3d(project, experimenter, num_cameras, working_directory)	Start a new 3D project
create_pretrained_project(path, task, videos, SUPERANIMAL_NAME, model_name, detector_name)	Create project from SuperAnimal pretrained model
add_new_videos(config_path, videos)	Add videos to existing project

Training & Evaluation

Function	Description
create_training_dataset(config_path, net_type, augmenter_type)	Prepare training data; net_type: resnet_50, resnet_101, mobilenet_v2_1.0, efficientnet-b0
train_network(config_path, maxiters, saveiters)	Train; key params: maxiters=100000, saveiters=5000
evaluate_network(config_path, plotting=True)	Evaluate on test set, produce metrics

Video Analysis

Function	Description
analyze_videos(config_path, videos, videotype, save_as_csv)	Predict poses for all frames in videos
create_labeled_video(config_path, videos, videotype, filtered)	Overlay predicted keypoints on video
video_inference_superanimal(videos, superanimal_name, ...)	Zero-shot inference with pretrained SuperAnimal models

Post-Processing

Function	Description
filterpredictions(config_path, video, ...)	Smooth predictions (ARIMA, median filtering)
analyzeskeleton(config_path, video, ...)	Compute bone lengths, joint angles from predictions
plot_trajectories(config_path, video, ...)	Plot body part trajectories over time

SuperAnimal Pretrained Models

Available models for zero-shot inference:

Model	Species	Body Parts
superanimal_topviewmouse	Top-view mouse (various strains)	21 keypoints
superanimal_quadruped	Quadrupeds (dog, horse, sheep, etc.)	39 keypoints
superanimal_face	Primate/human faces	54 keypoints
superanimal_full	Full body animals (topview mouse + quadruped)	Combined

Common Pitfalls

1. Wrong PyTorch version: Always install PyTorch BEFORE deeplabcut. Check pytorch.org for the correct CUDA version.
2. GPU out of memory: Reduce batch size (default 8 → 4 or 2) in pose_cfg.yaml.
3. Missing GUI dependencies: label_frames requires pip install "deeplabcut[gui]". On headless servers, use X11 forwarding or label locally.
4. Video codec issues: If create_labeled_video fails, try converting to .mp4 (H.264) or .avi first.
5. maDLC detector: Multi-animal mode requires a detection model (fasterrcnn_resnet50_fpn_v2 is the default). Single-animal can work with heatmap regression alone.
6. SuperAnimal scale: Adjust scale_list parameter for SuperAnimal — try [200, 300, 400] first; smaller animals may need higher values.