Home

Welcome to the Squamate Vertebra NSM Wiki!

This wiki provides comprehensive documentation for understanding, training, and using the SDF-based neural generative shape model built using >2000 Squamate vertebrae (many lizards and some snakes).

🦎 Quick Start

• Overview: High-level introduction to the project goals and architecture.
• Installation: Step‑by‑step instructions for setting up the environment.
• Data Preparation: How to format and preprocess input meshes and landmarks.
• Training Guide: Explanation of training stages, hyperparameters, and troubleshooting.
• Evaluation: Metrics, mesh quality checks, and tips for debugging failing outputs.
• Use cases: Shape completion (broken specimens), Classification (species and vertebral position), Landmark transfer, Latent space exploration

🦴 Use Cases

1. Shape Completion
   Fill in missing parts of an incomplete vertebrae - predict missing structures of a missing vertebrae by sampling points on intact surfaces. Guide to hyperparameters used for shape completion here.

2. Classification
   Classify an unknown vertebrae - predict the species and spinal position (Example: Tupinambis teguixin C3) based on the top 5 closest matches from our training dataset

3. Landmark transfer
   Transfer landmarks from mean/median specimen to other specimens to see how model is interpreting shape and position

4. Latent Space Exploration
   PCA, t‑SNE, UMAP, Isomap, visualization examples (video and grid generation), and interpretation methods.

🔮 Example Workflows

• Training from scratch
• Fine‑tuning on a new specimen set
• Generating meshes from random latent samples
• Projecting landmarks onto generated meshes
• Shape completion for partial vertebra specimens (ex: fossils)
• Classification (species and spinal position) for unknown vertebra specimens (ex: fossils)

---