Comprehensive Unsupervised Behavioral Explorer

From video to behaviour, end to end.

CUBE finds and quantifies animal behaviour in video. It combines pose tracking, unsupervised clustering, annotation and statistics in one open-source app. Clusters are found without manual labels, and no programming is needed.

Record Track Discover Annotate Analyse
Get started What is CUBE? Watch the overview
CUBE: Comprehensive Unsupervised Behavioral Explorer
The CUBE main window with step cards for DLC, pre-processing, clustering, annotation and analysis The CUBE main window in dark mode
Built on proven open methods
DeepLabCutB-SOiDUMAPHDBSCANHidden Markov modelsAnipose

How it works

Five steps from recording to results

Scroll through the pipeline. Each step saves its progress, so you can stop and pick up where you left off.

1

Record

Any camera, any rig. CUBE works with single-camera videos, dual cameras and synchronised multi-camera setups.

  • Multi-camera recordings synchronised, with fill-frame detection
  • Quality flags on every frame
Three synchronised camera views of a mouse with frame-by-frame quality flags
2

Track

DeepLabCut finds the body points in every frame. With calibrated cameras, CUBE triangulates them into a 3D skeleton.

  • SuperAnimal quadruped model, adapted to your videos with cleaned pseudo-labels
  • Likelihood and reprojection error checked
Keypoints on a mouse in one camera view next to the reconstructed 3D skeleton
3

Discover

Posture and movement features are embedded with UMAP and grouped by HDBSCAN into clusters, then smoothed over time with a hidden Markov model.

  • Cluster settings chosen by a validity sweep; stability across random seeds can be checked
  • Every cluster comes with its own posture sequence
UMAP embedding of movement windows with posture sketches for individual clusters
4

Annotate

Watch example clips of each cluster, give it a name once, and the name applies to every video.

  • Pose overlay on the example video
  • Optional guided labelling to check names against human labels
Three example clips named Grooming, Head dip and Rearing
5

Analyse

Ethograms, behaviour sequences, bout statistics, transitions and group comparisons are built in, and figures and tables can be exported.

  • Group comparisons and a multivariate group predictor
  • Everything exports as figures and tables
Ethogram, behaviour sequence, cluster posture and before-and-after plots

Flexible input

Works with many kinds of recordings

CUBE has been used with videos from several kinds of setups, from a tablet recording of a cage to classic mazes and calibrated multi-camera rigs.

  • iPad videos and other single-camera recordings
  • Dual cameras, and uncalibrated rigs of up to three cameras
  • Open field and elevated plus maze, viewed from above
  • Calibrated 3D rigs with four cameras
Examples of input videos: iPad, open field, elevated plus maze, dual camera and 3D multi-camera

Human in the loop

See what your names change

CUBE first finds clusters blind. When you name behaviours, it uses your labels to guide the clusters towards the behaviours you care about. Drag the handle to compare.

  • Unsupervised first: no categories decided in advance
  • Optional guided labelling, with performance reported on fully labelled videos
  • Names, clusters and statistics stay linked

How guided labelling works

Open field UMAP coloured by 24 blind CUBE clusters The same UMAP coloured by behaviour names from guided labelling Blind clustersWith your labels

Open field, top view, one shared UMAP. Left: 24 blind CUBE clusters. Middle: CUBE with guided labels, evaluated on held-out videos. Right: the human labels. Data: Sturman et al. (2020).

Why CUBE

Built-in checks along the way

Quality checks at each stage help you judge how far to rely on a result.

Quality checks

Confidence reports, frame-level quality flags, reprojection checks in 3D and seed-stability checks on clusters.

Temporal smoothing

A hidden Markov model smooths the classifier output over time, which reduces frame-to-frame flicker.

Guided labelling

Add your own labels to steer clusters towards the behaviours you define, and check the result on fully labelled videos.

Multi-camera 3D

Calibrated four-camera rigs are triangulated into a 3D skeleton. Uncalibrated rigs of up to three cameras fall back to a fused 2D view.

Aware of your arena

Trace regions and objects once to get paradigm metrics for open field, elevated plus maze, Y-maze, novel object and more.

Statistics built in

Ethograms, bout statistics, transitions, group comparisons and a group predictor with permutation testing, with no code to write.

Quick start

Up and running in minutes

CUBE runs on Windows 10 or 11. You do not need to install Python or write any code. An NVIDIA GPU is recommended for pose tracking and optional for analysis.

Full setup guide

# 1. Download the ZIP from GitHub and extract it
# 2. Double-click the installer (10 to 30 minutes, once)
install_cube.bat
# 3. Double-click the CUBE icon on your Desktop

Documentation

Learn CUBE your way

Pipeline guide

Every step explained, with settings, outputs and troubleshooting.

Read the guide →

Analyser and Group Predictor

Statistics, reclustering, transitions and group comparisons.

Open the reference →

Guided labelling

Use your own labels to steer clusters, and validate them.

See the method →

Behavioural paradigms

Open field, EPM, Y-maze, novel object and more.

Browse paradigms →

3D multi-camera

Calibration, triangulation and the fallback for uncalibrated rigs.

Set up 3D →

Gallery

Real outputs from open field and elevated plus maze analyses.

View the gallery →