AI3D-360: guardrail rail-vs-support split — real-segment validation

2026-08-15 · iolabs AI3D team · generated by Claude Fable 5 · all imagery: intensity rasters (A4_5 dataset policy)

Feature

The guardrail detector now separates the rail beam (B/W-profile band, 0.35–0.85 m) from its supporting posts. Each accepted rail run gets a corridor-scoped low-band evidence pass (0.10–0.50 m above road), an autocorrelation cadence estimate snapped to the Swiss/German catalog {1.33, 2.00, 4.00} m, and a guardrail_support companion instance — a new exported class in guardrails.json, its own XML <Feature>, the overlays (violet line, one dot per detected post), and point_masks.npz (support points re-labelled per point). Validated on real Abschnitt 4/5 segments; verdict: working, one real defect found on real data and fixed (station quantization, see Segment 118).

Rail instances carry post_spacing_m (snapped), post_spacing_raw_m, post_spacing_snapped, post_count, post_confidence and post_gap_spans (each span flagged occluded when it overlaps an occlusion interval already bridged on the rail — gaps mean unobserved-or-missing, never “defect”). Everything sits behind three config flags (enable_post_cadence, enable_support_class, enable_component_masks, all default on); with all three off, every output is byte-identical to the pre-feature detector.

Commit	What
`942f58e`	Post cadence + `guardrail_support` class (instances, masks, overlays, config)
`216e410`	Supports exported as their own XML `Type="guardrail_support"` Feature; evidence-dump lever
`9bf5baf`	Fix: continuous `RoadSpine.project_fine` station for the cadence passes (1.0 m grid artifact)

Segment 107 — clean 2.00 m cadence on all four rails

Segment 107 top-down overlay — Full tile. Green = `w_beam` rails (G00–G04), violet = the four `guardrail_support` companions (S05–S08) with one dot per detected post.

Segment 107 zoom on rails with post dots — Zoom: 25 posts over ~48 m on each rail — a 2.00 m train, snapped from raw estimates 1.997–2.001 m, confidences 0.52–1.0.

Segment 118 — the quantization artifact, before and after

The first real-segment run reported an unphysical 1.0 m spacing on two rails. The dumped evidence histograms showed 100% of the low-band mass in one residue class modulo 1.0 m: RoadSpine.project snaps stations to its ~1 m sample grid, which the 1 m-binned main pipeline never notices but which folds a 0.10 m cadence histogram onto the grid — and aliases genuine 1.33 m post trains to 4.0 m. Fixed by project_fine (along-tangent refinement), used only by the new passes so pre-feature outputs stay byte-identical.

Before — S06 fabricated 1.0 m posts

After — S06 genuine 1.33 m posts

Left rail (G02/S06, vegetation side): dot cadence drops from a fabricated 1.0 m (51 “posts”) to a genuine 1.33 m (38 posts, raw estimate 1.323–1.330 m) — the tighter ASTRA 11005 pitch, previously disguised as 4.0 m by aliasing. Median rails stay at 2.00 m.

Segment 118 full overlay after fix — Full tile after the fix: two 2.00 m rails, two 1.33 m rails, all snapped, zero gap spans.

Segment 135 — weak evidence is rejected, not invented

Segment 135 full overlay — Mixed segment (10 rails incl. far-field zone-recovered ones, noise walls in cyan). Only the two rails with solid low-band evidence get supports — both 2.00 m.

Segment 135 zoom on supports — Zoom on S11 (2.00 m, 48 m span) crossing the interchange; noise walls in cyan. Before the fix this segment reported nine cadences, seven of them the bogus 1.0 m; after the fix the weak far-field rails correctly get no cadence instead of a fabricated one.

Cadence results (after fix)

Segment	Parent rail	Spacing	Raw	Snapped	Posts	Confidence	Gap spans
107	G00	2.00 m	1.998	yes	25	1.000	0
107	G02	2.00 m	2.001	yes	25	0.983	0
107	G03	2.00 m	2.000	yes	24	0.878	1
107	G04	2.00 m	1.997	yes	25	0.521	0
118	G00	2.00 m	2.000	yes	26	0.939	0
118	G01	1.33 m	1.330	yes	38	0.860	0
118	G02	1.33 m	1.323	yes	38	0.739	0
118	G03	2.00 m	2.003	yes	25	0.994	0
135	G00	2.00 m	1.999	yes	25	0.984	0
135	G01	2.00 m	2.004	yes	26	0.514	0
135	G02–G10 (8 rails)	no cadence — weak/far-field low-band evidence, correctly rejected

Verification

pass 259 unit tests, incl. synthetic cadence (30% deleted posts), sub-mask precision on synthetic truth, station-quantization regressions, config/JSON parity.
pass Flags-off invariance on real segment 107: rail JSON exactly equal after stripping post_*; mask rows a strict subset with agreeing instance_id; XML equal after removing the support Feature.
pass point_masks.npz schema unchanged (supports ride the existing instance_type table; e.g. segment 107: 855,960 w_beam + 41,745 guardrail_support points).
pass XML: separate Type="guardrail_support" Feature with segment_NNN_guardrail_support_NN polylines through the Nexus exporter.

Notes & open items

tuning Height cut lines (0.35 / 0.50 m) are literature-derived config values; behaved well on 107/118/135, left untouched until a downstream consumer needs otherwise.
perf The evidence pass is O(runs × points): ~4 min on 4-rail segments, ~32 min on segment 135 (19 instances). Optimizable if it matters.
Support mean_height_m is the band midpoint (0.3), not measured; support height profiles are empty by design.
JSON point_count and mask row counts are different quantities (pre-existing); zone-recovered rails own no occupancy cells and thus no mask rows (pre-existing).
Branch t3code/separate-v-profile-support-classes, v0.3.1 + AI3D-360 commits; wrap-up (version bump, Nexus, push) pending.