Reviving the Statistical Shape Models

Bringing early ASM/AAM work back to life as a clean, modern, pip-installable library.

A long-running goal: take the Active Shape Model / Active Appearance Model work from years ago and rebuild it as a clean, documented, installable package.

Why bother in the deep-learning era

Statistical shape models are still unbeaten when you have few samples, need interpretable parameters, or must run on tiny compute. They pair well with modern landmark detectors as a refinement / regularization stage.

Planned scope

• PCA shape model construction from annotated landmarks
• Profile-based and texture-based fitting
• AAM appearance modeling
• Interop with common landmark formats

Status: scoping. Follow progress on the forum.