Reference surface with nearby Helios support
Measured on 10 August 2026 · 72-view orbit
Helios supported almost all CAD within 5 mm—and retained a wider off-surface tail.
A fixed-scale comparison of one Helios reconstruction with the exact PF-T31-00001059-C694BB02 corrected scan-visible CAD reference. The result separates CAD completeness from retained off-surface deviation and screens how far the 72-view orbit may be reduced.
Measured result
The 72-view reconstruction at a glance
Typical registered surface deviation
Width · depth · height bias
Image-space hypothesis; 3D rerun pending
Combined-system error, not scanner-only bias.
Residuals include the physical print, connector seating, Helios reconstruction, rigid registration, occlusion, and sparse sampling. No caliper or CMM measurement of this specimen was available.
Interactive registered geometry
Inspect where the scan departs from CAD
All 16,274 target-prism scan points are shown—no enrichment of the worst residuals. The CAD layer is a deterministic 8,000-point surface sample.
Preparing registered geometry
16,274 Helios points · 8,000 CAD points
Signed-normal summary statistics cover the 14,459 supported Helios points within 5 mm; signed colours on the retained tail are exploratory.
Choose Helios or an iPhone scan with the system menu. The point overlay is accompanied by keyboard-accessible camera presets, the full comparison table below, and downloadable metrics. Point-by-point picking requires a pointer.
Two directions, two questions
Completeness and surface deviation separate cleanly
Nearest-neighbour direction matters. Reporting only one direction would hide the core Helios trade-off.
Helios points→CAD
How far did reconstructed points depart from the model?
Median1.686 mm
p957.245 mm
RMSE3.167 mm
A wider tail remains around the base, tray boundary, thin edges, and difficult surfaces.
CAD→Helios points
How completely did the scan support the reference?
Median1.120 mm
p952.849 mm
≤ 5 mm99.93%
Nearly all sampled CAD has nearby Helios support, including much of the thin plant geometry.
Surface metrics by scope
All distances in millimetres| Scope | Scan→CAD median | Scan→CAD p95 | CAD→scan median | CAD→scan p95 | CAD ≤2 mm | CAD ≤5 mm |
|---|---|---|---|---|---|---|
| All target-prism points | 1.686 | 7.245 | 1.120 | 2.849 | 85.45% | 99.93% |
| Diagnostic: scan points ≤5 mm | 1.515 | 4.015 | 1.120 | 2.849 | 85.45% | 99.93% |
| Plant union | 1.383 | 4.054 | 1.312 | 3.402 | 79.27% | 100.00% |
The ≤5 mm row is a labelled diagnostic that removes the off-surface tail; it is not the primary result.
One CAD · four capture pipelines
Helios and three iPhones reveal different kinds of reconstruction error
Every column is rescored against the same 55,622-point FULL_BORE_v1 CAD sample. No single metric is allowed to stand in for “accuracy.”
Comparable scoring, not a controlled hardware shoot-out.
These are single complete capture-and-reconstruction runs with different trajectories, image counts, software, point densities, and registration routes. The iPhone 15 overlay was already plate-scaled and plant-aligned on device, then rescored with the common distance and extent definitions. Helios and the iPhone 12/17 runs retained fixed metric scale during their CAD registrations.
Highest sampled-CAD coverage within 2 mm.
Within 5 mm; iPhone 15 follows at 99.85%.
Best CAD→scan p95, narrowly ahead of iPhone 15.
Closest width/depth versus closest height.
Helios supported 42/44 leaflets; all runs supported 4 leaves and 5 fruits.
Versus 33:18 for the complete Helios orbit.
Same-CAD normalized comparison
Bold cells identify the strongest observed value for that row—not an overall winner.
| Metric | Helios | iPhone 17 Pro Max | iPhone 15 Pro Max | iPhone 12 Pro Max |
|---|---|---|---|---|
| Surface support and deviation | ||||
| CAD within 2 mm ↑ | 85.45% | 79.05% | 88.96% | 74.98% |
| CAD within 5 mm ↑ | 99.93% | 96.06% | 99.85% | 94.90% |
| Scan→CAD median ↓ | 1.686 mm | 0.841 mm | 1.190 mm | 0.739 mm |
| Scan→CAD p95 ↓ | 7.245 mm | 3.658 mm | 6.152 mm | 3.369 mm |
| CAD→scan p95 ↓ | 2.849 mm | 4.400 mm | 2.946 mm | 5.041 mm |
| Symmetric mean ↓ | 1.828 mm | 1.404 mm | 1.508 mm | 1.412 mm |
| Robust dimensions | ||||
| Width bias · |error| ↓ | +1.055 mm | +3.620 mm | +2.476 mm | +2.405 mm |
| Depth bias · |error| ↓ | +1.482 mm | +3.088 mm | +2.269 mm | +2.540 mm |
| Height bias · |error| ↓ | −2.943 mm | −3.679 mm | −0.884 mm | −4.742 mm |
| Mean absolute W/D/H bias ↓ | 1.827 mm | 3.462 mm | 1.876 mm | 3.229 mm |
| Reference-supported structure and acquisition | ||||
| Compound leaves | 4/4 | 4/4 high | 4/4 strong | 4/4 high |
| Fruits | 5/5 | 5/5 moderate | 5/5 strong | 5/5 moderate |
| Leaflets | 42/44 | 44/44 high | 44/44 strong | 44/44 · 39 high |
| Capture duration ↓ | 33:18 | 4:08 | 2:32 | 3:27 |
| Capture samples | 72 projection views | 139 RGB · 121 registered | 91 RGB · 72 exported poses | 120 RGB · 101 registered |
| Registered target points | 16,274 | 14,437 | 28,233 | 11,251 |
The iPhone 12/17 clouds are sparse, so their low Scan→CAD residuals partly reflect selective sampling of easier surfaces. CAD→scan and CAD-coverage rows expose what was missed. Capture durations exclude reconstruction and training time.
Conformity of retained points
Answers “how far are the reconstructed samples from the model?” It rewards clean points, but can look excellent when difficult or missing surfaces never produce points.
Completeness of the reconstruction
Answers “how far must each CAD sample travel to find reconstructed support?” It reveals absent leaves, gaps, and under-sampled structures.
Same scoring, two registration tiers
Helios and iPhone 12/17 retain fixed metric scale during CAD alignment. The iPhone 15 result keeps its package-authored plate scale and bounded plant alignment before common-definition rescoring.
Expected-organ support, not blind detection
Leaf, fruit, and leaflet counts test support at known CAD locations. They do not discover unexpected extra organs, and nearby parts can share points.
Device and geometry provenance. The iPhone 15 Pro Max model is verified in its HEIC metadata and the scored geometry is its Apple Object Capture photogrammetry mesh—not LiDAR or the package’s Gaussian splat. The iPhone 12/17 names come from their benchmark report; their package manifests say only “iPhone.” Those two columns score sparse Mac-side SfM clouds associated with this CAD through all four AprilTags, but no QR identity observation was packaged.
Shape and structure
Dimensions stayed close; thin leaflets were the main losses
Extents use registered p0.5–p99.5 surface bounds. Component counts are reference-supported—not blind detections.
Width151.055 mmCAD 150.000 mm
+1.055 mm +0.70%
Depth151.482 mmCAD 150.000 mm
+1.482 mm +0.99%
Height159.525 mmCAD 162.468 mm
−2.943 mm −1.81%
4/4compound leaves
5/5fruits
42/44leaflets
31 leaflets had strong support, 11 were moderate or partial, and 2 were insufficient.
Corner fiducials and CAD-conditioned volume Open advanced results
| Fiducial | Centre error | Local scale bias | Form RMSE | CAD→scan p95 |
|---|---|---|---|---|
| A · short cylinder | 0.274 mm | −2.21% | 0.956 mm | 1.865 mm |
| B · tall block | 0.493 mm | −1.63% | 0.988 mm | 1.735 mm |
| C · reinforced frustum | 0.217 mm | −0.52% | 0.848 mm | 1.439 mm |
| D · low dome | 0.109 mm | −1.43% | 0.903 mm | 1.536 mm |
The CAD-guided plant exterior-volume proxy was 37.716 cm³ versus a 37.555 cm³ CAD union (+0.429%). This is a model-assisted local-scale consistency check, not independent scan-only enclosed volume.
Projection-angle screen
Twenty-four views mark the image-space knee
The screen ranks angular redundancy. It identifies which subset to reconstruct next; it does not prove reduced-view 3D equivalence.
versus
Estimated sweep
11:06
22:12 saved versus 72 views
Balanced subset screenHeld-out composite coverage
Best phasePhase range
The 72-view value is not connected: with no held-out frames, its 100% score is definitional and not comparable to this curve.
Explore a balanced orbit24 selected views
Score80.03
Sweep11:06
Largest gap15°
Angles
5, 20, 35, 50, 65, 80, 95, 110, 125, 140, 155, 170, 185, 200, 215, 230, 245, 260, 275, 290, 305, 320, 335, 350
24
Recommended first rerun. Exact 15° spacing on the available 5° lattice; only 5.9% more held-out proxy error than 60 views.
20
Aggressive lower bound. Alternating 15°/20° gaps; 17.9% more held-out proxy error than 60 views.
Audit trail
What the benchmark does—and does not—claim
The primary surface result never rescales the scan. Every public download is sanitized and checksum-listed.
Registration and metric scale
The surrounding Helios tray is excluded with a known-target prism. A proper rigid transform is refined with 75%-trimmed ICP while holding scale at 1.000 mm/mm.
Eight 80% resamples changed translation by a median 0.106 mm and rotation by 0.064°. A free-scale diagnostic preferred +0.836%, but that rescaling is excluded from the primary bias.
The PCD itself does not declare units. Its native coordinates are consistent with millimetres under the known 150 mm target registration.
Reference provenance
The reference is the FULL_BORE_v1 corrected scan-visible CAD revision for PF-T31-00001059-C694BB02, derived from the target design and corrected assembly datums.
Despite the source filename’s “as-built” wording, this is not a CMM-measured physical mesh. Print warping, surface texture, and real connector seating error are not encoded.
Signed bias and dimensions
Signed residual follows the nearest local CAD triangle normal. Positive means outward. The all-supported median is +0.634 mm, but signs are less stable at thin leaves and edges where the nearest triangle can change.
Robust extents can contract when surfaces are missing. The plant-only horizontal envelope is therefore an occlusion diagnostic, not a direct material-size measurement.
Why 24 views are not yet validated
The finished PCD has no per-angle provenance. Removing points would test point decimation, not capture count. Each balanced image subset must be reconstructed independently in Helios.
The acceptance plan allows median worsening ≤0.5 mm, p95 worsening ≤1.0 mm, CAD-coverage drop ≤3 percentage points at 2 mm, additional height-bias magnitude ≤1.0 mm, and additional base width/depth bias ≤0.5 mm relative to 60 views, with all five fruits and four leaf modules supported.
Public result package
Take the metrics and geometry with you
Local paths and ambiguous secondary fields have been removed. Checksums cover every published result file.
JSONFour-system comparisonHelios and iPhone 17/15/12 metrics, registration tiers, recovery, and capture provenance
CSVComparison tableSame-CAD normalized values for analysis or plotting
3DiPhone 17 viewer points14,437 registered target points with absolute CAD residuals
3DiPhone 15 viewer points28,233 registered photogrammetry-mesh samples with CAD residuals
3DiPhone 12 viewer points11,251 registered target points with absolute CAD residuals
JSONBenchmark metricsSanitized fixed-scale results, traits, context, and angle screen
CSVMetric summaryUnique metric keys with explicit scopes and units
CSVProjection subsetsEvery tested view count and exact balanced angles
CSVHelios rerun planRecommended phases for the 3D confirmation experiment
PLYError-coloured Helios cloud16,274 registered points with absolute and signed residuals
PLYCAD comparison overlayRegistered scan and deterministic reference samples
JSONRegistration transformFixed-scale transform, bootstrap stability, and provenance
JSONFull angle analysisMethod, phase variants, descriptors, timing, and quality control
MDPackage notesProvenance, limitations, field definitions, and checksum instructions
SHAChecksumsSHA-256 for the published page data and downloads