STL/PLY 3D Tooth Library and Standardized Dental Rendering Pipeline

STL/PLY 3D Tooth Library, ExoCAD 441-Form Workflow and Standardized Dental Rendering Pipeline

1. Why 2D smile design previews fail inside 3D CAD environments

When flat 2D drawings or unconstrained generative inpainting outputs are handed over to a dental CAD technician working in ExoCAD or 3Shape, four systematic geometric failures occur:

1. Perspective distortion and focal-length drift: Close-up photographs captured with short focal lengths exaggerate the apparent width of central incisors while compressing canines and premolars along the buccal corridor. Applying a flat 2D golden proportion ruler (PMID:279670) without compensating for camera intrinsics produces unrealistically narrow lateral incisors and canines once wrapped onto a 3D dental arch.

2. Tooth-count and contour hallucination in unconstrained diffusion: Generative diffusion models synthesize plausible tooth textures but routinely violate anatomical topology. In internal benchmark logs (ACIK_ISLER.md:32), running an unconstrained image-generation pass caused the SAM (Segment Anything Model, arXiv:2304.02643) tooth count to jump from 5 to 9 teeth, pushed the 95th-percentile contour deviation to p95 = 48 pixels, and altered 38.8% of pixels outside the target mask. Consequently, unconstrained diffusion was removed from the delivery path in favor of deterministic 3D-to-2D compositing (DEVIR.md:124).

3. Z-axis depth and overjet ambiguity: A 2D overlay specifies vertical incisal length (Y) and horizontal width (X), but omits labio-palatal inclination (Z). Without a 3D transformation matrix (M_local_world) anchored to the intraoral scan, the CAD operator must guess whether to procline or retrocline the restoration, directly affecting lip support and phonetics.

4. Cervical root-extension clipping errors: Many CAD tooth libraries store meshes as open shells with long root extensions designed for gingival emergence boolean operations. Measuring crown height via a naive bounding box on teeth exceeding 12 mm captures only 55% of the true crown ratio and introduces false central-incisor asymmetry (ACIK_ISLER.md:27, issue A214).

2. The 441-form tooth library, 537 guide profiles, and 476/477 verified STL/PLY render pool

To replace subjective template picking with measured morphological matching, the library layer is structured into two parallel datasets: a 441 primary tooth-form catalog (backed by 441/441 base profiles and a total of 537 measured markdown guide files under ~/atlas/atlas_data/rehber_ic, median size 6,154 B ≈ 6 KB; DEVIR.md:100) and a 477-entry STL / PLY 3D rendering pool including repaired shell variants (ACIK_ISLER.md:23, arsiv/DEVIR_arsiv_2026-10-09.md:37).

Each tooth form is indexed as a verified 3D surface mesh with pre-computed geometric invariants:

• Central incisor width-to-height ratio (W/H): Evaluated against the anatomical crown reference range (0.78–0.85, PMID:12806322) established for unworn maxillary dentition.
• Disto-incisal radius and incisal embrasure geometry: Mapped across cuboidal (square), ovoid, and tapered archetypes using the dentogenic SPA (sex, personality, age) filter (DOI:10.1016/0022-3913(58)90043-x) and morphological archetype parameters.
• Surface macro- and micro-topography: Developmental lobes, marginal ridge prominence, and perikymata surface normals that govern specular reflection.

During the render-library verification closure (arsiv/DEVIR_arsiv_2026-10-09.md:37, commit 6fd6b12), 34 open-shell variants were repaired and added to the 442 base entries, resulting in 476/477 verified render packages (RENDERKUT_OK 18/18, BINEK_OK 71/71, ISARET_OK 29/29). Every package contains a locked triad—render_rgba, siluet_alfa (alpha silhouette), and dis_indeks (per-tooth index mask)—sealed with a single deterministic transformation hash (sha1).

3. Camera projection matrices, CIE LAB lighting, and subsurface scattering

Projecting a 3D STL/PLY arch into a 2D photograph without creating an artificial appearance requires headless Blender/OpenGL rendering governed by three physical constraints:

1. Photograph-camera projection (hekim_dizimi.py): Rather than rendering the tooth library with an arbitrary orthographic studio camera, the 3D mesh is rendered through the estimated camera projection matrix of the target photograph (DEVIR.md:122). Each render pass outputs seven synchronized layers: render_rgba, siluet_alfa, dis_indeks, derinlik (Z-depth), cfg, rapor.json, and the registered comparison overlay.

2. CIE LAB / CIEDE2000 color-space harmonization and incisal shadow modeling: Natural dental enamel exhibits anisotropic translucency over warmer dentin. In the compositing pipeline (DEVIR.md:100, commit 8feeb639), luminance (L*) and chromaticity (a*, b*) statistics are sampled in CIE LAB space from existing reference teeth; specular highlights and upper-lip incisal cast shadows (commit e8ffbcc9, DEVIR.md:115) are computed mechanically rather than hallucinated.

3. Pixel-level lip and gingiva lock (with Synthetic Benchmark Caveat): Pixels outside the dental silhouette—lips, gingival zeniths, facial hair, and mandibular teeth—are strictly immutable. In the locked synthetic benchmark (blog/en/face-to-tooth-cephalometric-lora.html:116), 4,686 surrounding pixels remained 100% identical across model comparisons while only the maxillary anterior six-tooth line changed. Mandatory Methodological Caveat (bot/olgular.json:O-008, O-009): The locked n = 32 evaluation set on which face-shape accuracy improved from 0.0% → 90.6% (29/32) and target tooth-line accuracy reached 100.0% (32/32) is a 100% rule-derived synthetic numerical test set, not a clinical patient photograph cohort.

4. Arch curvature, cervical boundary flags (0x800), and the p95 ≤ 0.2 mm gate

Once the 3D render is composited into the 2D Constrained Canvas, the Mechanical Fidelity Gate (kosucu/sadakat_kapisi.py --referans render, self-test selftest 41/41; end-to-end compositing runner ~/atlas/mockup2d/mockup_hatti.py --selftest 30/30; DEVIR.md:100,124) enforces three fail-closed checks:

• Exact tooth-count parity: The number of segmented teeth in the final composite must equal the tooth index mask (dis_indeks) of the 3D render.
• Zero out-of-mask leakage: Every pixel outside the constrained silhouette mask is restored from the original frame (out-of-mask difference = 0.000%).
• Contour deviation bound (p95 ≤ 0.2 mm): The 95th percentile of the Euclidean contour distance between the 3D render silhouette (siluet_alfa) and the final composite must not exceed 0.2 mm. On alpha-channel identity verification across three test packages, contour deviation measured p95 = 0.000 mm (3/3 PASSED, DEVIR.md:100).

Measurable Evidence Card (Live Library and Rendering Logs)

Every numerical metric on this page is read directly from live repository verification logs and locked datasets with data-kaynak attributes:

Frequently Asked Questions (FAQ)