Eye-Region Semantic Detection

One-click eye detection for figure models. Status: v1 written from verified sources. Usage guide: User Guide: Seed Tools.

Why eyes

Eyes are small, high-curvature and visually critical: no other region ruins a figure print as fast when segmentation gets it wrong. CYM detects them and treats them as protected seeds for everything downstream.

Two entry points

Command	Scope	Since
`detect_eye_regions(roi_faces)`	inside a face ROI; thresholds via `EyeParams` (parameterized in `6b63659`)	ROI era
`detect_eye_regions_auto()`	whole model, no ROI needed — template matching backend (`detect_eyes_global`, `TemplateParams`)	Stage 1, `8d0e063`

flowchart LR
    M["full model scan<br/>(auto) or ROI scan"] --> C["geometry cues /<br/>template match"]
    C --> F["filter & validate"]
    F --> E["eye-region labels<br/>protected in fuse / merge"]

Behaviour

Stale suggestions are cleared before a new detection runs — old recommendations can't contaminate the new result (3756321).
Lone eye-only grows are guarded — a seed grow that would end up eye-only without its body fails safe (3756321).
Protection — detected eye regions survive internal dihedral cuts (2fc242a) and fuse/merge passes.
Detection results land as regions/labels ready for seed grow & fuse — never as direct colour writes.

Code pointers

src-tauri/src/segment/template.rs — global template backend (Stage 1)
src-tauri/src/segment/eye.rs — eye primitives + EyeParams
src-tauri/src/commands/segment.rs — both IPC entries
src/components/SeedPanel.tsx — the "detect eyes" button

TODO

Template/cue pipeline detail (what the template matches, acceptance thresholds with provenance)
Failure gallery & manual recovery (re-seed by hand when detection misses)
ROI vs. auto accuracy comparison on the sample corpus