Feature/face anonymization

[Mah-diaa]

This work was developed as part of my thesis on facial anonymization for photogrammetry scans. I tried to abide as much as possible by the rules of contribution in the wiki. I will also attempt to explain how the more "vague" parts of this works. Hope this helps!

Summary

This PR adds a deletion-based face anonymization pipeline for photogrammetry scans used in fNIRS workflows. The pipeline takes a surface plus five anatomical landmarks (Nz, Iz, Cz, LPA, RPA), isolates the head from surrounding scan artifacts/body geometry, removes the facial/ear region while preserving landmarks and cap-relevant geometry, and exports an anonymized OBJ bundle.

What changed

• Added cedalion.geometry.photogrammetry.anonymization
• Added anonymize_scan(...) as the main public entry point
• Added helpers for axis normalization, head isolation, CTF alignment, cap-boundary detection, face-mask construction, vertex
  deletion, frame reversion, and anonymized OBJ export
• Added texture sanitization during export so removed face regions are not still present in the JPG texture
• Added examples/head_models/51_manual_5pt_anonymization.ipynb for the manual 5-landmark workflow
• Added synthetic unit tests in tests/test_anonymization.py

Design notes

• The workflow is landmark-based: the user provides Nz, Iz, Cz, LPA, and RPA.
• The scan is first aligned into a CTF-style anatomical frame, anonymized there, and then transformed back to the original digitized frame by default. I did this because I was unsure whether downstream cedalion workflows require the anonymized surface to keep the same frame convention as the original scan.
• Head isolation uses position-based seam merging for connectivity analysis. During experimentation I found that some photogrammetry heads are not represented as one clean connected component, but as several overlapping UV/geometry islands. This caused problems when selecting the largest connected region to remove scan artifacts.
• The anonymization mask tries to stop before it reaches the optode/cap region. It estimates where the front edge of the cap starts and uses that as an upper limit for face removal. (theres also a failsafe if the subject happens to wear a cap with no optodes)
• Several mask parameters are exposed on anonymize_scan(...), so the deletion region can be made more conservative or more aggressive if the default mask is too generous or too greedy for a given scan.
• The pipeline preserves UVs where possible and sanitizes texture output during saving.

Validation

Added synthetic mesh tests covering the exported API and the end-to-end anonymization path:

• missing landmark handling
• axis normalization
• CTF alignment
• frame reversion
• cap-boundary fallback
• face-mask semantics
• vertex deletion
• geometry-only save path
• full anonymization pipeline

Example Jupyter notebook

The example workflow is in:

examples/head_models/51_manual_5pt_anonymization.ipynb

It demonstrates the intended manual 5-landmark workflow:

• load an Einstar photogrammetry scan
• select Nz, Iz, Cz, LPA, and RPA
• run anonymize_scan(surface, landmarks)
• inspect the before/after result
• save the anonymized OBJ bundle with save_anonymized_scan(...)

Before running it, update the local input/output paths near the top of the notebook:

• SCANS_FOLDER: folder containing the subject scan folders
• SUBJECT_ID: subject folder or scan identifier to process
• OUT_DIR: destination folder for the anonymized OBJ/MTL/JPG output

Reviewer notes

• The implementation is isolated under src/cedalion/geometry/photogrammetry/anonymization/.
• The README currently contains thesis-related context because it is needed for my thesis submission. I expect to revise and clean it up once the PR is closer to being ready to merge.