Grass-overhang close-ups at the asphalt rim

2026-08-01 · AI3D-337 · raw lane-point renders (no voxelization), a1_b000 · camera on the road, 1.3 m inboard of the rim, eye 0.33 m above the asphalt plane, frame ~1.7 m wide · splats world-sized 4.5 cm with depth sorting · companion to the v2 veg-notch report

Why these exist. Since grass demonstrably overhangs the rim, top-down rasterizations (incl. z-diff) cannot see the true pavement edge there — the highest return wins the cell and the edge appears to retreat inward. These close-ups show the situation in 3D. Color code in the class views: grey = pavement returns (±8 cm of fitted asphalt plane), green→yellow = 0.08–0.45 m above plane (grass), and RED = pavement returns that have a grass return directly above them in the same 3 cm column — i.e. literal, per-point proof of asphalt under grass.

s049 — the genuine overhang asphalt under grass, confirmed

a1_b000 seg 49, left side, station 90.25. Grass tongue reaches ~0.27 m over the rim; the pavement returns underneath are as bright as the clean road surface (brightness ratio 0.99) — real asphalt, not soil.

True RGB, camera at knee height on the road. Grey pavement scan rows run under the green front; the guardrail is the pale band top-left. Scale bar 0.5 m.
Same camera, height-class. The red points along the interface are pavement WITH grass directly above (n=1381 in frame) — the overhang band, visible through gaps in the blades.
Cross-section, 0.2 m slab. Bottom panel: the grey pavement line continues left past v=0 and turns red under the green arc out to v≈−0.45 m. The blades hang 0.1–0.3 m above it — Miro's right-hand sketch, in data.
Top-down 2×2 m, 2.5 cm cells. Left: all returns — the grass line. Middle: pavement slice only — pavement extends past the visual grass line. Right: red cells (95) = both grass and pavement in the column = overhang footprint.
Wider context shot (2.8 m inboard, eye 0.55 m) for orientation.

s032 — grazing fringe ~0.1–0.15 m, dark under-surface

a1_b000 seg 32, right side, station 32.75. Overhang exists but is shallower, and the returns under the blades are 2.5× darker than clean road — shadowed asphalt or edge detritus; RGB can't distinguish.

Red overhang band clearly present along the rim, plus brown below-plane points at right where the verge dips.
True RGB of the same view.
Cross-section: the bright asphalt band starts at v≈−0.05; red under-grass points hug the rim only. The tall grass mass sits on a verge bank slightly above the plane.

s062 — grazing fringe ~0.15 m, dark under-surface

a1_b000 seg 62, right side, station 58.50. Same pattern as s032: convincing green-over-grey in perspective, but the surface beneath the blades is dark (brightness ratio 0.32) — not provably clean asphalt.

True RGB: grass mass riding onto the outer pavement rows; guardrail above.
Height-class view of the same camera.
Cross-section slab at the densest overhang evidence.

Takeaway for the detector

Where grass overhangs (s049-style), a top-down raster puts the edge at the grass line, not the pavement rim — the error equals the overhang width (0.1–0.3 m here). Recovering the true rim needs the pavement-height slice (middle top-down panel) or per-column two-return logic (red cells): both are computable from the same npz points that feed rasterization. Full stats per site in output/vegnotch/cand_<id>/closeup_stats.json; renderer output/vegnotch/scripts/pc_closeup.py.