explore-behavior 2.0.2

Automated exploration behavior analysis for object recognition and location tests
CLIP-based classification with an iterative labeling loop — no GPU required.

Overview

EXPLORE analyzes rodent object exploration from overhead video recordings. It uses CLIP (ViT-B/32) to classify frames and a lightweight logistic-regression head that you train interactively in a few minutes.

Core workflow (browser GUI):

1. Add videos and name your project
2. Draw bounding boxes around each object on a reference frame
3. Verify that boxes are correctly re-localized across all videos
4. Sample 1-minute windows, correct proximity-based labels, train the classifier
5. Run the full analysis — annotated videos, exploration CSVs, trajectory plots

Installation

# 1. Install PyTorch (CPU)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

# 2. Install EXPLORE
pip install explore-behavior

For GPU (CUDA 12):

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install explore-behavior

Quick start

GUI (recommended)

explore gui

The browser app opens automatically. Work through the six tabs in order:

Session state is saved automatically at each step — use Load project on Tab 1 to resume from where you left off.

Output files

DI and RI

Discrimination Index and Recognition Index are computed for every pair of objects when Compute DI / RI is enabled. With objects A and B:

DI_A_vs_B = (t_A − t_B) / (t_A + t_B)
RI_A_vs_B = t_A / (t_A + t_B)

Values are based on total session exploration time (not per-bin), which avoids instability from empty time bins.

Development

git clone https://github.com/victorjonathanibanez/EXPLORE
cd EXPLORE

pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
pip install -e ".[dev]"

# Run tests (no GPU, no model download)
pytest -m "not slow and not gpu"

# Lint + format
ruff check src/ tests/
ruff format src/ tests/

Contributing

Pull requests welcome. Areas where contributions are most valuable:

• Pose estimation integration for more robust tracking
• Better object assignment when multiple objects are very close together
• Windows-specific testing

Please open an issue before submitting large changes.

License

MIT — see LICENSE.

Citation

If you use EXPLORE in your research, please cite:

Ibañez, V., Bohlen, L., Manuella, F. et al. EXPLORE: a novel deep learning-based analysis method for exploration behaviour in object recognition tests. Sci Rep 13, 4249 (2023). https://doi.org/10.1038/s41598-023-31094-w

Contact

Victor Ibañez — victor.ibanez@uzh.ch
Anna-Sophia Wahl — AnnaSophia.Wahl@med.uni-muenchen.de