Software · DeepLabCut

DeepLabCut™: a software package for animal pose estimation

An efficient method for 2D and 3D markerless pose estimation, based on transfer learning with deep neural networks. It matches human labeling accuracy with minimal training data, typically 50–200 frames, and works across species and behaviors. It is open source, fast and robust, and supports 3D pose and multiple animals.

Cheetah
Mice, multi-animal
Mouse reaching
Horse
Mouse whiskers
Pupil tracking
Racehorse
Mouse on a wheel
Mouse following an odor trail
Fruit fly
Human reaching, in 3D

Behavior, measured across species with DeepLabCut

Key publications

The science behind DeepLabCut.

Primary research

    Reviews, benchmarks & more

      Where do you start?

      Free, open source and easy to install.

      Follow the installation instructions, then use the graphical interface and the step-by-step user guide.

      DeepLabCut works across species: mice, flies, cheetahs, humans, octopus and more
      DeepLabCut Model Zoo

      Pretrained models, no training required.

      Use models already trained on a wide range of animals and scenarios to analyze your videos directly, or fine-tune them with about 10× less data than the original DeepLabCut weights. Read the SuperAnimal paper (Ye et al., Nature Communications 2024).

      SuperAnimal-Quadruped model card
      SuperAnimal

      SuperAnimal-Quadruped

      Trained on over 40,000 images of four-legged animals, from mice, rats, horses, dogs and cats to elephants and gazelles. 39 keypoints.

      SuperAnimal-TopViewMouse model card
      SuperAnimal

      SuperAnimal-TopViewMouse

      Trained on over 5,000 mice across diverse lab settings. Mostly C57 mice, with some CD1 (white) mice. 26 keypoints.

      State-of-the-art models, built in

      BUCTD pose estimation in crowded scenes of people, marmosets and fish
      ICCV 2023

      BUCTD

      Bottom-up conditioned top-down pose estimation for crowded scenes. A bottom-up model proposes each individual's pose, and a conditional top-down model refines it, so overlapping animals or people are kept apart.

      CVPR 2026

      FMPose3D

      Monocular 3D pose estimation via flow matching: it lifts 2D keypoints from a single image into 3D. Pretrained for humans and for quadrupeds, with the animal pipeline built on SuperAnimal.

      Contribute

      Help us make the Model Zoo better with more data, more models and more results. Label images in the DeepLabCut labeling web app, email modelzoo@deeplabcut.org, or propose a model through this form.

      How to cite a model

      Please cite the paper listed with each model. For our own unpublished models (non-commercial use only), please cite Ye et al., 2024.

      Thank you to our Model Zoo contributors. Jim McBurney-Lin (UC Riverside), Claire Witham (MRC Harwell), Rollyn Labuguen & Jumpei Matsumoto (University of Toyama) contributed models. Sam Golden, Jacob Dahan, Katherine Shaw, Nick Hollon, Loukia Katsouri, Lauren (Nikki) Beloate, Leandro Aluisio Scholz, Diego Giraldo, Alice E Kane, Nina Harano, Vasyl Mykytiuk, Khaterah Kohneshin, Michelle Joyce, Alejandra Del Castillo, Tara O'Driscoll and Alice Tang contributed labeled data.

      Contact

      For general inquiries or media: admin@deeplabcut.org.

      Please do not reuse logos or images from this page without permission, except for research use with a citation of the source.