Skip to content
AnimalPose logo

Animal Pose

AnimalPose finds rodents in video, tracks each animal's body through 25 keypoints, and scores behavior.

Table of contents

Structure

Multiview

Camera model
and
drift correction

Fuse views
into 3D

Video

Detect

Pose

Track

Refine

Behavior

Assay
(experiment rules)

Stage What happens
Detect finds every animal in each frame (a box and a center)
Pose places 25 keypoints on each animal, from nose to tail tip
Track keeps each animal's identity from frame to frame
Refine smooths each track and fills short gaps with a temporal network
Behavior labels each frame with a behavior class
Assay applies an experiment's rules (zones, entries, contact, …) and writes per-session measures
Multiview matches the views of a rig, corrects small camera movements and fuses the keypoints into 3D

Tracked two mice demo

Original Video Credit (1)

  1. CRIM13, Burgos-Artizzu, Dollár, Lin, Anderson and Perona, "Social behavior recognition in continuous video", CVPR 2012 (data.caltech.edu, CC BY), session 110308_A22_Block14; cut to 3.4 s and overlaid by AnimalPose.*

A general-purpose pretrained network is included, and you can train your own networks for new species or behaviors. See Training.

Usage Options

  1. Desktop app: a user-friendly GUI for recording and analyzing experiments without code.
  2. Command line (rust): a single Rust program for batch processing, servers and scripts.
  3. Python package: a full implementation for training, evaluation and research.

Packages

Each GitHub release has:

File Contents
animalpose-gui-v<version>-<platform>-<system architecture>.* the desktop app
animalpose-v<version>-<platform>.* the command line
animalpose-models-v<version>.zip the general models and sample profiles, for the command line
animalpose-<version>-py3-none-any.whl the Python package (also on PyPI)
SHA256SUMS checksums of every file

You may need FFmpeg. See Quick start.

Requirements

MacOS 14 or later on Apple Silicon; Intel Macs not supported.

Windows 10 or Windows 11; x86-64.

Linux x86-64 or ARM64; with glibc 2.35+ , OpenGL.

Python 3.10 to 3.13; PyTorch 2.3+, NumPy 2; FFmpeg (ffmpeg and ffprobe).

License

GPL-3.0-only for the code.

All Rights Reserved for the models, datasets, and pretrained networks. You may not use them for commercial purposes without a separate license.