AI3D-373 — vegetation split: pictures for the open questions
Four delivered LAS variants of the AI3D-373 vegetation split, rendered as oblique close-ups so the shape and height of the green stuff is visible — a top-down raster cannot tell a 0.3 m grass strip from a 1.5 m box hedge. Every question below is a decision I cannot make from the numbers alone.
A = segment 085, from_ground=TRUE · B = segment 085, from_ground=FALSE (the current default) · C = segment 003, from_ground=TRUE · D = segment 060, from_ground=FALSE. A and B are the same point cloud with the same knobs (exg 0.10); the only difference is whether green "tablecloth" ground rows are allowed to become low_vegetation. Flip between the A and B picture of a pair and the only thing that moves is yellow-green vs brown.
How to read the renders
- Left panel = class colours (registry palette), right panel = the raw sensor RGB of the very same pixels. The legend under the left panel counts the points of each class inside the window.
- Projection is orthographic, so metres-per-pixel is constant across the frame. The white L in the corner is 1 m across and 1 m up; the yellow ruler through the centre is ticked every metre off the ground plane. A hedge that reaches the "2 m" tick really is 2 m tall.
- The small circular inset is a 0.10 m/px top-down thumbnail of the same window with a red arrow showing the camera direction — use it to place the oblique on the map.
- Default camera is 20° above the horizon, azimuth 45° and 225°. Where that looked along the carriageway and the guardrail became a wall, I re-shot across the road (az 135/315) or from higher up; the title bar of every image states the actual az/elev.
Class colours
low_vegetation 3medium_vegetation 4tree 5ground 2asphaltguardrailvehiclesolid_line
Q1 — hedge, wall, or trees?
Decision: across the three probe segments the verticalsigns hedge rule flipped exactly one cluster from
tree to medium_vegetation: this one, 8.4 × 7.4 m on the ground, 4.6 m tall. Is that a hedge (rule right), a wall/noise (rule right for the wrong reason), or a group of small trees (rule wrong)? If it is the third, the whole hedge rule earns its keep on one false positive.
tree column standing 0→4.5 m against the ruler with dark-green medium_vegetation wrapped round its base — no flat face, no straight top line, so this reads as vegetation rather than a wall. Only 2,939 points in the whole 8 m disc (930 medium / 822 tree / 733 ground / 297 low) — this is a far, thin part of the scan, which is why the splats are enlarged here.
tree cluster 45 m up the road that the hedge rule did not flip: 8.4 × 7.9 m footprint, 3.1 m crown. Look at: tree 5,249 / low 3,828 / medium 3,220 — a real broad crown sitting on a low-veg bank, with orange guardrail at the left edge for scale.
Q2 — keep the detector trees?
Decision: the tree detector keeps things the vegetation split would otherwise have banded as medium or low. Two cases: a 4.8 × 4.8 m, 4.5 m-tall "tree", and the verge behind the guardrail where the QC pass saw 1–2 m scrub labelled
tree interleaved with medium and low. Keep them as trees, call them hedge, or call them grass?
from_ground.
low_vegetation carpet running up the bank, dark-green medium patches inside it, and mint tree blobs at the far edge that are only 1–3 m tall by the ruler. asphalt 365,751 / guardrail 50,997 / ground 33,715 / low 29,092 / tree 22,263 / medium 7,240.
from_ground does.
Q3 — is the 0.5 m low/medium cut right?
Decision: everything under
low_max_m=0.5 is low_vegetation, everything from 0.5 to medium_max_m=2.0 is medium_vegetation. These are the two boundary spots the auto-picker chose on 085 (cells where both classes live within 1 m of each other), plus four auto-picked spots on a middle segment. Does the yellow-green/dark-green line fall where you would draw it by eye?085 boundary spots, A vs B
from_ground takes away roughly two thirds of both vegetation classes, so this spot is mostly tablecloth ground with sparse growth on top.
Segment 060 (D, from_ground=FALSE) — four auto-picked spots
medium mat covering the embankment from 0 to ~2 m with essentially no low (medium 15,774 vs low 161). Hedge or just tall grass on a slope?
Q4 — tablecloth ground → low vegetation, and the median
Decision:
from_ground decides whether green rows that the ground fitter has already claimed (the "tablecloth") may be re-labelled low_vegetation. A (TRUE) paints the verges yellow-green; B (FALSE, current default) leaves them brown. Second half of the question: the median between the carriageways, where a grass ribbon sits between two guardrails and a false vehicle ribbon covers it — how far does the asphalt guard reach in there?The median, seen from above the guardrail
guardrail + pink vehicle, with only a thin yellow-green fringe. asphalt 1,275,918 / guardrail 285,852 / ground 120,104 / vehicle 66,094 / low 13,340 / medium 812 / tree 0.
from_ground barely matters — the grass there is lost to guardrail/vehicle, not to ground. That is the asphalt-guard reach question, not the from_ground question.
vehicle ribbon running the full length of the median at 0–1 m — a moving-object artefact lying exactly where the grass is. Nothing in the median is ever offered to the vegetation stage.
60 × 60 m overview of segment 085, A vs B
low_vegetation apron between the guardrail and the tree line. ground 752,277 / tree 314,776 / low 112,637 / medium 24,317.
from_ground=TRUE moves ~99k points from ground to low and ~11k to medium in this one 60 × 60 m window — the RGB half is the tiebreaker on whether that apron is grass or bare shoulder.