App for dinosaur footprint analysis via disentangled variational autoencoder. Youtube DinoTracker mini-doc (English): https://youtu.be/ZbGp8tIOfqw

The project and the Dinotracker app have been covered by international media outlets, including:

• The Conversation –
  Identifying dinosaurs from their footprints is difficult, but AI can help

• Deutschlandfunk (German) –
  Versteinerte Spuren: KI-App klassifiziert Fußabdrücke von Dinosauriern

• The Guardian –
  Scientists launch AI Dinotracker app that identifies dinosaur footprints

• BBC Newsround –
  Meet Dinotracker: the AI app that identifies dinosaur footprints

• Reuters –
  New AI method helps identify which dinosaur made which footprints

This repository provides the app described in the following paper:
G. Hartmann, T. Blakesley, P.E. dePolo, & S.L. Brusatte,
Identifying variation in dinosaur footprints and classifying problematic specimens via unbiased unsupervised machine learning,
Proc. Natl. Acad. Sci. U.S.A. 123 (5) e2527222122,
https://doi.org/10.1073/pnas.2527222122 (2026).

• data/images_compressed.npz → Raw data used for training and testing (1 MB)
• data/names.npy → Names of the images (300 KB)
• data/Tone_logo_small.png → The app's logo (42 KB)
• data/tracks.xlsx → Overview of the used tracks (88 KB)
• models/model_BETA15_BIG_3k_shuffle_epoch1000.pth → beta-VAE model (2.5 MB)
• models/mu.npy → Encoding of the images (78 KB)
• src/Create_training_data.py → Creation of training data (9 KB)
• src/Training.py → Training the beta-VAE (9 KB)
• src/DinoTracker.py → The app for footprint analysis (31 KB)
• docs/Installation_guide.pdf → See below "Environment Setup" (965 KB)

This application was tested with:

• Python 3.13.5
• PyTorch 2.9.0
• PyQt6 6.9.1
• NumPy 2.2.6
• pyqtgraph 0.13.7
• PIL 11.3.0
• h5py 3.14.0

To make this application accessible to researchers from different backgrounds, we provide a step‑by‑step illustrated guide for setting up the Python environment. This guide was created based on feedback from early testers and is available here: /docs/Installation_guide.pdf.

python src/DinoTracker.py

We warmly welcome feedback, feature requests, and ideas for improving this application.
If you have suggestions or encounter any issues, please contact:

📧 gregor.hartmann@helmholtz-berlin.de

To further improve the training of our neural network, we are actively seeking additional dinosaur footprint data.

• Silhouettes
• Photographs
• Other relevant image material

If you have access to such data and are willing to share, your contribution would be highly valuable to expanding our dataset and enhancing the accuracy of footprint analysis.

All contributions will be acknowledged, and shared data will be used strictly for research and development purposes.

The training of the network was performed on an Apple Mac Studio equipped with:

• M2 Ultra
• 24‑core CPU
• 76‑core GPU
• 192 GB unified memory

The code is optimized to take advantage of the large unified memory available on this system.
For users without access to such hardware, an alternative data loader is provided to enable training on machines with more limited resources.

On Apple Silicon systems, the recommended backend is:

• device = torch.device("mps")

On other platforms, replace "mps" with "cpu" or "cuda" depending on availability:

• device = torch.device("cpu") # for CPU-only systems
• device = torch.device("cuda") # for NVIDIA GPU systems

Note: This section is relevant only for training the network from scratch. For the DinoTracker app, we provide a fully trained network and recommend to use cpu inference, which is the default setting.

• Python (PSF License)
• PyTorch (BSD-style)
• PyQt6 (GPL-3.0)
• NumPy (BSD)
• PyQtGraph (MIT)
• Pillow (PIL License)
• h5py (BSD)

This application is released under GPL-3.0 to comply with PyQt6 licensing. All other dependencies are permissive and compatible with GPL-3.0.