# DeepLabCut: markerless pose estimation for animal behavior

> DeepLabCut is a Python toolbox for markerless pose estimation of user-defined features with deep learning, for all animals including humans. Version 3 runs on PyTorch with pretrained SuperAnimal models, offers a GUI alongside the API, and is developed by the Mathis labs under an LGPL licence.

**DeepLabCut/DeepLabCut** — Official implementation of DeepLabCut: Markerless pose estimation of user-defined features with deep learning for all animals incl. humans

- Repository: https://github.com/DeepLabCut/DeepLabCut
- Website: http://deeplabcut.org
- Stars: 5,780 · Forks: 1,792
- Language: Python
- License: LGPL-3.0
- Published: 2026-09-10 · Updated: 2026-09-10 · Language: en
- Canonical page: https://hysenlabs.com/projects/deeplabcut-deeplabcut

## Label what you can see, track anything

DeepLabCut's framing is refreshingly literal: as long as you can see and label what you want to track, you can use the toolbox, because it is animal and object agnostic. The target audience is dual, computer vision researchers who want state-of-the-art models and frameworks, and life scientists who get what the README calls best-guess defaults, served through both a GUI and an API. Development comes from the Mathis labs, credited in setup.py as A. and M. Mathis Labs, and the workflow is documented in a Nature Protocols paper with a step-by-step project management guide. The repository is under active current development, with the last push on 2026-09-29, the same day as this writing.

## PyTorch first, TensorFlow on a deprecation clock

The quick start installs the engine before the toolbox. PyTorch comes first, plain or CUDA-flavored:

```bash
pip install torch torchvision
```

```bash
conda install pytorch cudatoolkit=11.3 -c pytorch
```

Then the toolbox itself, with the GUI or headless:

```bash
pip install --pre  "deeplabcut[gui]"
```

Python 3.10 or newer is the supported floor, with conda environment files in the repository for guidance. A TensorFlow engine still exists through the tf extras, including CUDA-targeted variants like deeplabcut[tf-cu11] and deeplabcut[tf-cu12], but the README states plainly that the TensorFlow backend is aimed for deprecation in version 3.2, release date to be determined. New projects should read that as a one-way sign toward PyTorch, and existing TF pipelines should budget for the migration.

## SuperAnimal: pretrained quadrupeds and top-view mice

Not every user trains from scratch. Two foundation pretrained models ship with the project, SuperAnimal-Quadruped and SuperAnimal-TopViewMouse, and their evaluation methodology is documented in detail: the quadruped model was trained on its own dataset with the AP-10K collection held out for out-of-domain testing, the top-view model with the DLC-openfield dataset held out, per a 2024 paper in Nature Communications by Ye et al. Model weights are published through Hugging Face for reproducing the numbers, while the full models are distributed through the project's own dlclibrary. For a lab starting pose estimation, the practical effect is that a pretrained quadruped model can be evaluated on new footage before anyone labels a single frame, which changes the cost of finding out whether the tool fits the data.

## Benchmark tables you can read before choosing

The README carries concrete accuracy numbers, which most research toolboxes hide in papers. On the held-out AP-10K benchmark, top_down_resnet_50 reaches 54.9 mAP and top_down_resnet_101 reaches 55.9, while top_down_hrnet_w48 reaches 55.3 and top_down_hrnet_w32 52.5. On lab mouse open-field data the same models score far higher, 93.5, 94.1, 93.8 and 92.4 mAP respectively, the difference between wild photographs and a fixed camera in a controlled box. The spread between the two benchmarks is the useful lesson: architecture choice matters less than domain match, and the project directs users to an architectures guide in the documentation before selecting among the many models available in DeepLabCut 3.0. Tooling for training custom models with various backbones is provided alongside.

## Colab, Jupyter, Docker and a demo dataset

The examples directory is a runnable curriculum. A Colab notebook walks through pose tracking of a single mouse in an open field, and further demos cover running DeepLabCut from scratch on your own data. Demo datasets are checked in directly, Reaching-Mackenzie-2018-08-30 and openfield-Pranav-2018-10-30, so an installation can be verified against known data before real footage is touched. The test scripts enumerate the capability matrix by filename: single and multi-animal workflows in both PyTorch and TensorFlow, SuperAnimal inference, adaptation and transfer learning, and a transreid script for identity tracking. Docker support and Google Colab round out the deployment options, which matters for shared lab machines where installing CUDA stacks by hand is the usual failure point.

## A dependency list that mirrors a working lab

The runtime dependencies read like an inventory of a functioning vision lab. torch at 2 or newer with torchvision and timm for backbones, einops, and albumentations pinned at <=1.4.3 for augmentation, opencv-python-headless for image work, numba and filterpy for filtering, pycocotools for evaluation, and huggingface-hub for model distribution. The pandas dependency is pinned to >=2.2,<3 with the hdf5 and performance extras, and the pyproject comment links the pandas 3.0 migration to a tracked issue, the kind of note that tells you maintenance is managed rather than reactive. Pydantic v2 holds configuration. The licence is LGPL-3.0-or-later, weak copyleft, meaning modifications to the library itself must be shared while applications using it as a library face fewer obligations, with the LICENSE and NOTICE.yml files as the reference.

## Version 3.0.x, a forum and a residency

Release history shows a project mid-rewrite settling down: v3.0.0, the major version, arrived 2026-05-22, followed by v3.0.1 on 2026-07-28 and v3.0.2 on 2026-09-23, with a changelog directory tracking the details. Community infrastructure spans the scientific image analysis forum at image.sc under a deeplabcut tag, a Gitter channel, and a presence on X, plus contribution and code-of-conduct documents. Funding includes a Chan Zuckerberg Initiative Essential Open Source Software grant, and there is an AI Residency program alongside free workshop materials for the course-inclined. An AGENTS.md in the tree means agent-assisted contribution is anticipated here too. What ties it together is longevity: this is a maintained research tool with a decade-shaped support structure, not a paper's companion repository.

## Conclusion

Use DeepLabCut when you need markerless tracking of animal behavior and want a toolbox with a GUI, worked examples and pretrained models rather than a from-scratch pipeline. Consider SLEAP, the other well-known open source animal pose tracking project, when comparing research-group lineages before committing. Verify first that your Python is 3.10 or newer, decide PyTorch versus the TensorFlow engine knowing the TF backend is slated for deprecation in version 3.2, and check the architectures guide before choosing a model, since the benchmark tables reward that half hour of reading.

## FAQ

### What is DeepLabCut used for?

DeepLabCut is used for markerless pose estimation: tracking user-defined body parts of animals, including humans, in videos without physical markers. It serves computer vision researchers and life scientists studying animal behavior.

### How do you install DeepLabCut?

Install PyTorch first, for example pip install torch torchvision, then run pip install --pre "deeplabcut[gui]" for the GUI version or pip install --pre "deeplabcut" for the headless one. Python 3.10 or newer is required, and full instructions are in the installation documentation.

### Is DeepLabCut free?

Yes, it is free and open source under the LGPL-3.0-or-later licence. Development has been supported in part by a Chan Zuckerberg Initiative Essential Open Source Software grant.

### Is DeepLabCut machine learning?

Yes. DeepLabCut performs pose estimation with deep learning, supporting custom model training with various high-performance backbones as well as pretrained SuperAnimal foundation models for quadrupeds and top-view mice.

## Sources

- [DeepLabCut/DeepLabCut on GitHub](https://github.com/DeepLabCut/DeepLabCut)
- [License: LGPL-3.0](https://github.com/DeepLabCut/DeepLabCut/blob/main/LICENSE)
- [Project website](http://deeplabcut.org)
- [README](https://github.com/DeepLabCut/DeepLabCut/blob/main/README.md)
- [Releases](https://github.com/DeepLabCut/DeepLabCut/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/deeplabcut-deeplabcut