Create and tune scan-to-CAD methodologies with progressively more complicated models #60

New issue

Closed

opened 2026-08-13 00:54:55 +00:00 by Mark · 0 comments

Mark commented

2026-08-13 00:54:55 +00:00

Owner

Goal

Develop a systematic methodology for converting 3D scans to CAD models, tuning the pipeline at each level of complexity before escalating.

Methodology Tiers

Tier 1 — Simple primitives

Basic geometric shapes (cubes, cylinders, spheres, cones)
Single-material, no overhangs, known topology
Question: Can we fit primitives from point clouds directly? What tolerance is achievable?

Tier 2 — Extruded profiles

2D profiles with uniform extrusion (L-brackets, flanges, brackets)
Planar faces, right angles, through-holes
Question: Can we detect the extrusion axis, extract the profile, and reconstruct the solid? Does the L-bracket coverage baseline (issue #55/#58) lock in at this tier?

Tier 3 — Revolved/lofted bodies

Shafts, pulleys, handles with symmetry
Requires detecting axis of revolution or loft paths
Question: Can we segment rotational vs extruded features from the same scan?

Tier 4 — Compound assemblies

Multi-part scans where parts touch or interlock
Segmentation into individual components, then reconstructing each
Question: What's the reliable separation distance? Can we handle press-fit interfaces?

Tier 5 — Organic/freeform surfaces

Automotive body panels, ergonomic handles, sculpted forms
May require surfacing (NURBS) vs solid modeling
Question: What's the threshold where B-rep can't keep up and surfacing is required?

Workflow Per Tier

For each tier:

Generate or capture representative test scans (synthetic + real)
Run existing pipeline — see where it breaks
Tune parameters — segmentation thresholds, fitting tolerances, simplification
Lock in methodology — document the approach that works
Escalate — the breaking failures at this tier become the scope for the next tier

Deliverables

Per-tier methodology doc (what works, what doesn't, parameters)
Test corpus of representative scans (link to data files)
Benchmark table: tier × metric (coverage, deviation, time)
Decision tree: given a scan → which tier approach applies

#55 L-bracket coverage fixit
#58 L-bracket coverage still failing

## Goal Develop a systematic methodology for converting 3D scans to CAD models, tuning the pipeline at each level of complexity before escalating. ## Methodology Tiers ### Tier 1 — Simple primitives - Basic geometric shapes (cubes, cylinders, spheres, cones) - Single-material, no overhangs, known topology - **Question:** Can we fit primitives from point clouds directly? What tolerance is achievable? ### Tier 2 — Extruded profiles - 2D profiles with uniform extrusion (L-brackets, flanges, brackets) - Planar faces, right angles, through-holes - **Question:** Can we detect the extrusion axis, extract the profile, and reconstruct the solid? Does the L-bracket coverage baseline (issue #55/#58) lock in at this tier? ### Tier 3 — Revolved/lofted bodies - Shafts, pulleys, handles with symmetry - Requires detecting axis of revolution or loft paths - **Question:** Can we segment rotational vs extruded features from the same scan? ### Tier 4 — Compound assemblies - Multi-part scans where parts touch or interlock - Segmentation into individual components, then reconstructing each - **Question:** What's the reliable separation distance? Can we handle press-fit interfaces? ### Tier 5 — Organic/freeform surfaces - Automotive body panels, ergonomic handles, sculpted forms - May require surfacing (NURBS) vs solid modeling - **Question:** What's the threshold where B-rep can't keep up and surfacing is required? ## Workflow Per Tier For each tier: 1. **Generate or capture** representative test scans (synthetic + real) 2. **Run existing pipeline** — see where it breaks 3. **Tune parameters** — segmentation thresholds, fitting tolerances, simplification 4. **Lock in methodology** — document the approach that works 5. **Escalate** — the breaking failures at this tier become the scope for the next tier ## Deliverables - Per-tier methodology doc (what works, what doesn't, parameters) - Test corpus of representative scans (link to data files) - Benchmark table: tier × metric (coverage, deviation, time) - Decision tree: given a scan → which tier approach applies ## Related - #55 L-bracket coverage fixit - #58 L-bracket coverage still failing