Greenery & vertical markings

Two objects, two very different difficulty levels — the easy one first

Posts and signs are a near-solved problem with plain geometry. Greenery is the one object among all four where machine learning genuinely helps — and even there, a lightweight method comes first.

Topic 03 · the easy one

Vertical markings — posts & signs High confidence Low effort

Finding upright objects is one of the most mature corners of this whole field, and the recipe barely changes across a decade: remove the ground, then keep whatever is thin, upright and tall. That's it.

Posts found on rural highway⁶98%

Older classic method⁷86%

Telling the type apart⁸96%

The best highway result is 98% found, across 28 km of real road, with no training data.⁶ There's even a free, open-source extractor that runs on a plain CPU.⁹ And to tell a sign from a lamppost from a sapling, one team computed just four numbers per object (how tall, and how "line-like / flat / blobby" it is) and fed them to a tiny classifier you could train over lunch — 96% correct.⁸

➜ Recommendation Geometric pole extraction (thin + upright + tall), then a tiny four-number classifier for the type. Re-check the type thresholds on our own highways, since the 96% figure comes from a single city study.

Topic 02 · the hard one

Greenery — trees & bushes Medium confidence Medium effort

Everything that made the others easy — regularity, standard shapes, running parallel to the road — a tree has none of it. A tree is chaos: points scattered in every direction. That's actually the key to detecting it.

We measure, around each point, whether its neighbours are arranged like a line, a flat surface, or a scattered blob. Vegetation is the blob; man-made objects are lines or planes.

Feeding a bundle of these shape measurements to a light classifier reaches about 95%,¹⁰ and the same idea separates vegetation at over 98% in natural terrain — it's the method behind a tool (CANUPO) we already have on our desks.¹²

✗ Why it needs a bundle, not one clever trick Two "one simple rule" shortcuts for vegetation were rejected by fact-checking: that a single density number separates poles from trees (0 of 3 checkers),⁶ and that one shape ratio splits trees from signs (rejected).⁸ A leafy tree and a big sign can look alike on any single measure — you need several together.

If the light classifier isn't enough, a small 3D neural network is the next step — but with a warning worth remembering: in a proper benchmark, no single model won on all datasets (three datasets, three different winners). The leaderboard you read online will not transfer to our scans. The good news: the model the authors recommend for limited hardware is small enough to fit our graphics card comfortably.¹¹

➜ Recommendation Start with shape-statistics + a light classifier (no GPU). Only move to a small 3D neural net if that falls short — and pick the model by testing on our data, not by leaderboard. Greenery is the one class where we should budget for making some labelled examples.