Seed-Based Segmentation: Grow & Fuse

Seed-driven refinement for figure-type models. Status: v1 written from verified sources. Usage guide: User Guide: Seed Tools.

The flow

flowchart TD
    S["seed recommendation<br/>planar · multiview · cross-section · saliency"] --> G["manual grow<br/>seed_grow(seeds, barrier_deg, optimizer)"]
    S --> F["auto fuse<br/>fuse_segmentation (async)"]
    F --> T["tiny-region merge<br/>aggressive absorb + live size breakdown"]
    G --> RG["regions (labels only)"]
    T --> RG
    EY["eye-region detection"] -. protected from cuts/merges .-> RG

Seeded watershed growth

seed_grow takes placed seeds (SeedInput: snapped 3D point + hit face), a barrier angle and an optimizer flag. Growth spreads from each seed across the face graph and stops where the normal deviation exceeds the barrier — the seeded watershed model dates from iteration 50. The optimizer post-pass tightens the grown boundaries.

TODO: exact watershed scoring function and the barrier measurement definition (dihedral vs. normal-to-seed).

Seed recommendation

recommend_seeds(count, curvature, concavity) — saliency-weighted suggestions; the two weights are the panel's curvature / concavity sliders.
Structured detectors, each with its own parameters (see IPC Reference for signatures):
- Planar regions — detect_planar_regions(angle_threshold_deg, dist_thr_factor, min_region_faces)
- Multiview — detect_multiview_regions(view_count, angle_threshold_deg, min_region_faces, match_threshold)
- Cross-section — detect_cross_section_features(planes_per_axis, feature_threshold)

Fuse & tiny-region merge

fuse_segmentation (async) consolidates over-seeded output:

Aggressive tiny-region merging — micro-regions are absorbed so the result stays printable; a live breakdown of region sizes streams to the status area while it runs (iteration 80).
Cut threshold 2 — the merged-graph guard was hardened against pathological inputs (b543ede, "bulletproof tiny-region merge").
Reverse-split — a merge that would destroy structure can be reversed; the fuse-debug tooling from iteration 80's journal came out of exactly that failure mode.

Eye-region protection

Once eye regions are detected they are protected: internal dihedral cuts skip them (2fc242a) and fuse/merge passes keep them intact — a merged-away eye is a ruined figure print.

Invariants

All of the above propagate labels only. Overwriting face_colors with palette colours during fuse/grow is a known violation being tracked (tracker §8, pending the 3MF colour-strategy ruling) — see Fill Routing.

09-连续平面与边界-算法融合研究