Written and maintained by CASRAI Editorial Board
Last updated
Short answer: DeepLabCut is an open-source software toolbox for markerless pose estimation: it locates user-defined body parts of animals (or any labelled object) in video frames using deep neural networks, so behaviour can be quantified without attaching physical markers. It tracks position; it does not by itself decide what behaviour is occurring.
Definition and scope
The project’s documentation describes DeepLabCut as “a toolbox for state-of-the-art markerless pose estimation of animals performing various behaviors,” and states that it is animal- and object-agnostic: anything you can label can be tracked. The foundational paper (Mathis et al., Nature Neuroscience 21(9):1281-1289, 2018, DOI 10.1038/s41593-018-0209-y) frames the problem this way: physical markers are intrusive, and their number and location must be decided in advance. The method instead uses transfer learning with deep neural networks, and the authors report that it tracks body parts across multiple species and reaches accuracy comparable to human labelling when only a small number of frames (about 200) are labelled.
Two boundaries are worth stating plainly. First, pose estimation outputs coordinates of body parts over time; classifying those coordinates into behaviours (grooming, rearing, social interaction) is a separate downstream analysis. Second, the toolbox is software for analysing video you have already recorded under an approved protocol; it does not replace ethical review of the animal work itself (see IACUC and the entry on research animals).
How it is used in practice
The official user guide organises the workflow into four stages:
- Set up a new project, which creates the project folder and configuration.
- Label frames in the labeling GUI, marking the body parts you want tracked.
- Train and evaluate a network on the labelled frames.
- Analyze video with the trained network to obtain body-part positions.
Per the current documentation, DeepLabCut supports Python 3.10 and later, a GPU is optional and selected by choosing the matching CUDA version, PyTorch is the default backend in version 3.0, and the team aims to deprecate the TensorFlow backend in version 3.2. Multi-animal pose estimation is supported, with animal re-identification introduced in version 2.2.
Related concepts
Multi-animal tracking
Videos containing several animals can be analysed, with re-identification to keep individuals distinct over time (introduced in 2.2).
SuperAnimal pretrained models
The Model Zoo offers SuperAnimal foundation models, for example SuperAnimal-Quadruped and SuperAnimal-TopViewMouse, which can be applied to new videos without training from scratch. They are described in a 2024 Nature Communications paper (DOI 10.1038/s41467-024-48792-2). The documentation states that the SuperAnimal models are for research use and non-commercial only.
Markers versus markerless
| Aspect | Physical markers | Markerless (DeepLabCut) |
|---|---|---|
| Animal contact | Markers are attached; the paper calls them intrusive | Video only |
| Tracked points | Number and location set a priori | Body parts defined by the user when labelling |
| Effort | Hardware setup | Labelling a small set of frames (about 200 reported) and training a network |
Source for every row: Mathis et al. 2018 abstract and the DeepLabCut documentation.
Where it matters for labs
Because results depend on the labelled frames, the trained network and the software environment, a lab that wants its tracking to be reproducible needs to keep records tied to each documented stage. The following checklist is an editorial mapping of general reproducible research practices onto the toolbox’s documented stages, not a standard issued by the DeepLabCut project:
- Project setup: keep the project configuration file with the data, as part of the dataset’s metadata.
- Labelling: retain the labelled frames and a note of which body parts were defined.
- Training and evaluation: record the backend (PyTorch or TensorFlow), the DeepLabCut version and the Python version, since the supported backend and Python requirements have changed between releases.
- Pretrained models: if a SuperAnimal model was used, record which one and note that its terms are research-use, non-commercial only.
- Citation: cite the software and papers you used; see the FAIR4RS and software citation principles.
Licensing differs by component: the software is primarily licensed under the GNU LGPL v3.0, while the documentation states that the project’s artwork and logo are copyrighted and not covered by the software license.
Common misunderstandings
- “It classifies behaviour.” It estimates body-part positions. Behaviour labels come from analysis built on those positions.
- “It only works on mice.” The documentation describes it as animal- and object-agnostic; the paper reports tracking across multiple species.
- “It needs thousands of labelled frames.” The 2018 paper reports human-comparable accuracy with about 200 labelled frames in its tested settings; results for your own videos depend on your data.
- “A GPU is mandatory.” The documentation describes GPU use as optional.
FAQ
What is DeepLabCut used for?
Tracking user-defined body parts in video of animals or other objects without physical markers, typically for quantifying movement and behaviour.
Is DeepLabCut free?
The software is primarily licensed under the GNU LGPL v3.0. SuperAnimal pretrained models are stated to be for research and non-commercial use only.
Which Python version does it need?
The current documentation (read 2026-11-08) states that only Python 3.10 and later are supported.
Sources
- DeepLabCut documentation, deeplabcut.github.io/DeepLabCut (read 2026-11-08).
- Mathis A. et al. (2018). DeepLabCut: markerless pose estimation of user-defined body parts with deep learning. Nature Neuroscience 21(9):1281-1289. doi:10.1038/s41593-018-0209-y (PMID 30127430; record read via Europe PMC 2026-11-08).
- SuperAnimal pretrained pose estimation models for behavioral analysis. Nature Communications (2024). doi:10.1038/s41467-024-48792-2.
