Explore free-movement behavioral patterns
Discover patterns from posture and temporal features, then interpret clusters using behavioral sequences and source video.
XUANHENG · BEHAVIORAL RESEARCHDemo available
DIVE
Behavior-pattern discovery in freely moving mice
Train or reuse self-supervised representation models from experimental video, discover behavioral units through unsupervised clustering, and analyze pattern distributions, transitions and experimental-group differences.
- Self-supervised training and checkpoint resumption
- Unsupervised discovery of behavioral units
- Video-level and experimental-group comparisons
Enlarge imageRESEARCH APPLICATIONS
Discover behavioral patterns and structure from free-movement videos
Analysis inputs Single-mouse free-movement videos organized by experimental group
Exploratory behavioral phenotyping
Describe phenotypes using pattern duration, event counts and transitions to formulate further research hypotheses.
Compare behavioral structure across groups
Compare pattern distributions and transitions between groups and examine statistical tables. Differences alone do not establish causality.
MODEL TRAINING AND REUSE
Train representation models on your experimental data
Reuse an existing model or train on your own posture data. The learned representation supports feature extraction and unsupervised clustering.
- 01
Build a posture library and evaluate models
Organize posture reference images and assess how well default or custom models distinguish the current data.
- 02
Build training data
Generate globally augmented and locally masked training data from posture images.
- 03
Train, monitor and resume
Monitor progress, loss and validation curves, with pause, continuation and checkpoint resumption.
- 04
Save models for analysis
Save the best model and training records for reuse in feature extraction, temporal fusion and pattern analysis.
Representation-model evaluation
Feature example
Enlarge imageTraining parameters, evaluation, and saved outputs
- Training parameters
- Epochs, batch size, learning rate, weight decay, warm-up epochs, local masking ratio and number of frozen layers.
- Training evaluation
- Progress, elapsed time, remaining time, training loss and validation AUC, with records organized by parameter group.
- Models and outputs
- Best representation model, latest checkpoint, training history and curve PDFs.
CORE CAPABILITIES
Analyze, review, and organize results
Videos organized by experimental group
Link single-mouse videos to groups, configure sampling and outputs, and retain video-level and group-level relationships.
Segmentation and posture-image review
Inspect missing detections after instance segmentation, select targets and crop posture images to reduce background interference.
Self-supervised model training and resumption
Build augmented data, configure and monitor training, save the best model and parameter-group records, and resume from checkpoints.
Temporal fusion and unsupervised clustering
Fuse temporal features and use visualization and silhouette scores to support cluster-count selection, then inspect labels and behavioral sequences.
Behavioral distributions and transitions
Inspect pattern duration, event counts and transitions at video and experimental-group levels.
Group comparison and figure export
Perform between-group statistics and export data tables, PNG and PDF while retaining pattern definitions and analysis conditions.
Software in action
Actual interfaces from the DIVE user manual; original screenshots retain their Chinese labels.
Posture reference library and representation learning
Feature example
Enlarge imageTwo- and three-dimensional feature projections
Feature example
Enlarge imageRESEARCH OUTPUTS
Results and research interpretation
Figures from the DIVE user manual. Interpret pattern meanings alongside the original video.
Behavioral event transitions
Result example
Enlarge imageVideo-level behavioral distributions
Result example
Enlarge imageANALYSIS WORKFLOW
From data preparation to research outputs
- 01
Create projects and experimental groups
Organize single-mouse videos by experimental condition and check sampling settings and output locations.
- 02
Inspect segmentation and posture images
Check missing detections and target extent, ensuring posture images represent the research subject.
- 03
Train or reuse a representation model
Evaluate suitability, perform self-supervised training or resume it as needed, and use the model to extract and fuse temporal features.
- 04
Select clusters and interpret patterns
Inspect clustering through visualization, silhouette scores and behavioral sequences, and interpret the meaning of each pattern.
- 05
Compare and export research material
Review distributions, transitions and group differences, then export statistical data and result figures.
DELIVERABLES
Your analysis outputs
- Trained models and training records
- Posture images and temporal features
- Cluster labels and behavioral profiles
- Distributions and transition relationships
- Statistical tables, PNG and PDF
BEFORE ANALYSIS
What to confirm before you start
- Designed for single-mouse free-movement video, not multi-mouse identity tracking. Video and target conditions require verification.
- Unsupervised patterns require interpretation against source video and do not automatically correspond to predefined behavior names.
- Three-dimensional dimensionality-reduction plots represent feature space, not three-dimensional body reconstruction. Group statistics must be interpreted in the context of the experimental design.
FREQUENTLY ASKED QUESTIONS
Before you begin
How does DIVE differ from predefined behavior recognition?
DIVE explores patterns without predefined labels rather than directly producing named behavior categories. Researchers interpret cluster categories before analyzing their distributions and transitions.
Can I train a model using my own experimental data?
Yes. Build a posture library and augmented data, choose initial weights, and use self-supervised training with pause and checkpoint resumption. This trains a representation model, not a classifier for predefined behavior names.
How should cluster counts and group differences be checked?
Interpret clusters using feature visualizations, silhouette scores, behavioral sequences and video, then inspect distributions, transitions and statistical tables. Categories and differences alone do not establish a behavioral diagnosis or causality.
Find the right approach for your research
Tell us your experiment, recording angle, and target measurements to arrange a DIVE demo and sample assessment.
For research use only. Figure captions describe their sources and scope. Examples do not guarantee accuracy, diagnostic suitability, or research conclusions.