Plan 050: Landmarks and what LoD2 leaves out — measured in DOM1, named by OSM and Wikidata
Executor instructions: Read fully first. The code is built; what is open is the look on a real GPU (
bun run shots --headed, full profile) and baking the landmarks on every committed site. Update the status row indocs/plans/README.mdwhen a phase lands.Builds on plan 034 (the scan's small structures appended to the city mesh: the same append path,
sourcecolumn and footprint rules), plan 027 (OSM facts per LoD2 object,osm_buildings.py, theflagscolumn) and ADR 0041 (bridges measured in DOM1, typed by Wikidata; its fetch-time Wikidata cache is the pattern here).Drift check (run first):
git log --oneline -5 -- pipeline/bake/structures.py pipeline/bake/landmarks.py pipeline/bake/osm_buildings.py scripts/bake-city-mesh.ts
Status
- Priority: P2 (a city's silhouette and its names; every site)
- Effort: M (two bakes S–M, the look S, the HUD S)
- Risk: MED — false positives from the surface model (cranes, pylons, crowns); a loud colour or sheen breaking the clay
- Planned at: 2026-09-28
- Status: BUILT (2026-09-28), look unverified on a real GPU. Baked
on the seven committed sites in step order (
fetchfor the Wikidata landmarks, thenosm-buildings,landmarks,structures): 386 columns, 920 missing buildings, 508 relief slabs on 74 landmark objects, 310 landmarks (Dresden 180 / 370 / 196 / 142; Unna 5 / 32 / 34 / 6). The relief as slabs was replaced by a height field on 2026-09-30 (Findings); re-bakestructureson every site — a file baked before it carries slabs without agrid, which build nothing.
Why this matters
The maintainer asked (in German) for four things: search the gaps between the surface model and LoD2; give buildings an appearance from their attributes, in the app's style; derive geometry where it can be derived; a landmark list per city — all of it repeatable for every future city. Three buildings make the case:
- Unna, the Lindenbrauerei chimney — the "Fibonacci-Reihe" (Mario
Merz's neon numbers on a 48 m brewery chimney) is the town's
silhouette and is not in LoD2. OSM maps it as
man_made=chimney; DOM1 sees 48 m. - Hamburg, the Elbphilharmonie — LoD2 draws a flat block at 96.3 m. DOM1 reaches ≈ 109 m over the wave roof. Wikidata knows it with 41 sitelinks, and OSM tags its outline with the item.
- Dresden, the Congress Center — a glass facade under a curved roof;
LoD2 has flat parts 13–27 m high and, like every LoD2 object, no
material or colour. OSM's
building:materialandbuilding:colourare where that lives.
LoD2 has neither slim structures nor materials nor colours, and no dataset says which buildings a city is known by.
Design
Phase 0 — measurement (done)
gap = DOM − max(DGM, LoD2 roof) over the tile and a 60 m margin at 1 m
(the neighbours' DGM, LoD2 and DOM mosaicked, skyview.Roofs burning the
roofs). Findings:
- Tall slim gaps are common and mostly not buildings: tree crowns, power pylons and, on the flight day, cranes. Leipzig's centre: dozens of 50–95 m spikes with no mapped structure; Hamburg: 90–98 m spikes beside a building finished in 2023.
- Where OSM maps a
man_madechimney, tower or mast, the gap is there and its height is right; its width is not — a 1 m surface model sees a slim shaft too thin (the Lindenbrauerei's measured 1.7 m). - On landmarks, the excess over the LoD2 roof is the roof's real form (the Elbphilharmonie); on ordinary roofs it is antennas, dormers, lift housings and trees.
Gate: add nothing from the surface model alone. Passed as a design rule, not a threshold.
Phase 1 — structures beyond LoD2 (structures step)
pipeline/bake/structures.py → data/<site>/dlm/structures_<tile>.geojson
(ODbL; empty for a site without DOM).
- Columns: OSM
man_made=chimney/tower/mast/communications_tower/ water_tower/lighthouse, point or outline. Kept when the gap within 6 m (2 m of an outline) reaches 5 m and the structure 8 m. Axis at the gap's peak; foot at the lowest ground; height from the DOM, or OSM'sheightwhen the DOM undershoots it by > 25 % (a lattice mast); radius from the outline, OSMdiameter, else the gap blob low down; a chimney at least h/24; clamped per kind (chimney 0.8–6 m, mast 0.3–2.5 m, …) with a per-kind taper. - Missing buildings: an OSM building outline ≥ 40 m² (no roof, carport, construction site, ruin, underground or negative layer), LoD2 under < 20 % of it, ≥ 60 % of its cells ≥ 2.5 m above ground; a flat roof at the 60th percentile of the measured height.
- Landmark roof relief (after phase 3's
landmarks): per LoD2 object of a landmark, where the DOM stands ≥ 3 m above that object's highest LoD2 roof on ≥ 60 m² (and ≥ 2 % of its footprint), the excess as a height field per connected patch (≥ 6 m²): its outline and the 1 m grid of heights above that roof (grid:x,y= north-west corner,res,cols,rows,z, −1 outside), lightly smoothed (half the cell, half a Gaussian σ = 1 cell normalised over the patch),kind: relief,of= the object. (First built as slabs in ≥ 2 m bands, at most 8 — see Findings.) - Towers LoD2 draws: a
tower,lighthouseorwater_tower(BUILT_TOWERS) whose mapped foot lies under a LoD2 roof is skipped — LoD2 draws it already. Masts and chimneys on roofs stay. - A structure is written by the tile owning its anchor.
Build: lib/city/structures.ts lathes a column from its kind's ring
profile (16 sides), extrudes a building and builds a relief's height
field as a surface (reliefMesh: each corner the mean of the patch cells
around it, walls down to the roof along the patch's edge, no floor);
appendGapStructures (scripts/bake-city-mesh.ts) appends them to the
city mesh like the scan sheds, source = 2. A column is not a Building (the HUD's count);
chimneys, towers and lighthouses take the brick palette with the roof
tint = the wall's. A relief copies its host's row (tint, roof, flags,
root: demolish takes it along) and has no footprint.
Phase 2 — surfaces from attributes (osm-buildings, extended)
building:material / building:facade:material / facade:material
(normalised to glass, metal, brick, stone, concrete, wood, plaster),
building:colour, roof:colour (CSS name or hex → #rrggbb) on building
and building:part outlines covering ≥ 50 % of an object → material,
colour, roof_colour in osmbuild_<tile>.json; parts after buildings,
so a part's own tags win; a part inherits its root's (inheritedLook).
Build (lib/city/building-tint.ts): a mapped colour keeps its hue, its
saturation capped at 0.45 and its lightness held in [0.45, 0.9] (walls)
or [0.3, 0.8] (roofs) — a red wall becomes a clay terracotta. Walls: OSM
colour > material family > use family. Roofs: DOP > OSM roof:colour >
palette. Glass and metal set flags 4 and 8 (lib/city/city-mesh.ts);
the clay shader (osmColour, clayGlow) turns them a little cooler and smoother,
glass with a Fresnel sky sheen on the rim slider, dimmed at night.
Veto: no window grid, no panes, no texture.
Phase 3 — landmarks (landmarks step + fetch)
- Fetch (
landmarks.fetch_wikidata, called bybun run fetchfor every provider): SPARQL per tile,wikibase:box, instance of a subclass of Q811979 (architectural structure), ≥ 2 sitelinks, top 80 by sitelinks; label (de), position, height P2048, materials P186, building = subclass of Q41176. A box that times out (504) is re-asked in quarters, two levels deep. →data/_raw/<provider>/wikidata/landmarks_<tile>.json(CC0). - Bake (
pipeline/bake/landmarks.py): an item's LoD2 objects are those an OSM outline taggedwikidata=<Q>covers by half, else (a building) the LoD2 building under its point, root and parts. Items that match nothing drawn (districts, streets, harbours) are dropped; ≤ 12 per tile, most sitelinks first; Wikidata's material read as a wall material. →data/<site>/dlm/landmarks_<tile>.json. - Build:
withLandmarks(prepare-data.ts) setsOBJECT_FLAG_LANDMARK(16) and fills the material where OSM names none;siteLandmarksputs the site's twelve most notable into the tileset'sextras.landmarks. - HUD: the Erkunden tab lists them as Wahrzeichen chips under the
viewpoints; a click glides to
landmarkVantage(viaoverlook: from the south-south-west, altitude clamp(h + 50, 60, 300) m, pitch −22°, fov 55°).
Docs
Ledger (the five entries and four 🗃️ rows), data-flow.md,
rendering.md codebook, data-pipeline.md, portability.md, the guide
(data sources, using the viewer, how it works; en + de), ADR 0036.
Phases
- Measurement (above). Done.
structuresbake + tests (a chimney beside a hall; a landmark's roof as a height field, a spire to its tip; an ordinary roof's antennas make none; a church tower under its LoD2 roof makes no column) + the append. Built.- Looks from OSM + tests (colours, materials). Built.
- Landmarks: fetch, bake, build, HUD + tests (the timed-out box in quarters; a landmark point beside its building). Built.
- Open: finish baking every committed site (
osm-buildings,landmarks,structures, in that order), then plates on a real GPU: Unna's chimney and the Lindenbrauerei; the Elbphilharmonie's relief from the vantage and from the Elbe; the Congress Center's glass at dusk and noon; a mapped red and a mapped green facade beside hashed neighbours; a Leipzig centre view (no crane); Unna's Stadtkirche spire and Meißen's cathedral towers (one tower each, not two).
STOP conditions
- A column or block appears where nothing stands on the DOP: tighten the gap rule before adding a kind.
- A mapped colour reads louder than a DOP roof beside it: lower the saturation cap, not the clamp per colour.
- The glass sheen reads as a highlight or a reflection: halve it before tuning anything else; it must stay clay.
- A relief reads as a stack of plates from the ground: it did (the slabs), and the height field replaced them. If the height field reads as a lumpy blanket, smooth less or drop it — never an invented smooth roof.
Findings (the bake on the seven sites)
-
Relief measured cell by cell against the LoD2 roof was wrong: a courtyard's trees over its low roof and every roof step a metre off counted (the Museum für Kunst und Gewerbe in Hamburg got 99 slabs). Measured against the object's highest LoD2 roof (its 98th percentile) they stay below; the 15 % share gate then missed the Elbphilharmonie, whose LoD2 roof already slopes to 97 m and whose crests reach 12 m above it on 173 m² (3 % of the footprint) — so ≥ 60 m² and ≥ 2 %.
-
What the relief now finds is what LoD2 cuts short: the Heinrich-Hertz-Turm's top (to 271 m), the spires of Leipzig's Peterskirche and Neues Rathaus, the City-Hochhaus's mast, the Elbphilharmonie's crests, and Unna's Katharinen-Hospital, which LoD2 carries at 9.3 m on a 2017 floor plan while the surface model measures 16–22 m more over 2 100 m².
-
Wikidata's point is often beside the building (the Lindenbrauerei and the Zentrum für Internationale Lichtkunst in it: a courtyard): matched to the nearest LoD2 building within 25 m.
-
Place outlines carry
wikidatatoo: HafenCity and Wachwitz were matched through them to a quarter of LoD2 objects; only outlines of buildings, building parts andman_madestructures count now. -
A dense box times out at the endpoint (504 on a Hamburg and a München tile): it is asked again in quarters.
-
The chimney floor: the Lindenbrauerei's 48 m shaft measured 1.7 m across in the 1 m surface model; with r ≥ h/24 it is 4 m at the foot.
-
Stacked slabs made a truncated spire a stepped pyramid (Unna's Stadtkirche: LoD2 stops at 137 m, the surface model reaches ≈ 178 m) and the Elbphilharmonie's crests a flight of terraces — this plan's own STOP. The relief is now the measured height field on the 1 m grid, lightly smoothed, built as one surface: the spire rises to a point, the crests roll (2026-09-30).
-
The relief hovered at the ridge (Unna's Evangelische Stadtkirche, 2026-10-01): LoD2 draws the church as one gabled block (ridge ≈ 137 m, eaves ≈ 122 m); the west tower (walls ≈ 147 m, spire ≈ 187 m) is only in DOM1. The relief is measured above the highest roof, so its walls ended at the ridge's height and stood in the air beside the slope — and where the laser saw through the tower's sound openings to the ground, those cells tore "legs" down into it. Now the height field carries a
floorper cell — the lowest LoD2 roof in the 3 × 3 around it (roof cells only), sunk 0.3 m — andreliefMeshdrops each edge wall to it; DOM cells more than 1 m below the LoD2 roof under them are voids, filled from the nearest measured cell before the excess is taken (fill_voids). Unna's tower now stands on the nave's slope, its walls down to ≈ 126 m (floor −10.1 m under the 136.3 m roof). Only Unna'sstructuresis re-baked so far; a file withoutfloorkeeps the old walls. -
Towers were drawn twice: a church's or castle's tower is in LoD2, and the column search lathed a second one inside it — Meißen's cathedral towers (81 / 83 m), Grimma's churches, 13 of Munich's 14 towers. A tower, lighthouse or water tower whose foot is under a LoD2 roof is now skipped (
BUILT_TOWERS). -
Hamburg's ground was 0 m everywhere: the committed DGM tiles were all zeros (an old mosaic bug), so every building — the Elbphilharmonie among them — floated 5–10 m above the ground. Re-fetched; the fetch now refuses a height mosaic whose 1–99 % spread is under 0.5 m, a pipeline test holds every committed DGM to ≥ 2 m, and every Hamburg bake ran again on the real terrain.
-
Underground rails at street level (Munich's Marienplatz and Odeonsplatz looked broken): the Basis-DLM files U-Bahn and S-Bahn trunk lines as ordinary rail. A stretch within 2 m of a DLM tunnel (
BWF=1870) for > 15 m is cut (Munich: 23 km U-Bahn, the 4.2 km trunk line, 0 m of mainline); a 6 m reach had taken the Hauptbahnhof's surface tracks. DLM trams are left to the OSM tram layer (Munich's rail layer ≈ 84 → 22 km), OSM platforms and tram ways below ground are skipped, and a tram through a paved square gets the street bed (Munich's ballast ≈ 13 → 3 km, Dresden's 23.8 → 11.6 km). Leipzig's City-Tunnel left the Markt (815 m of rail there → 0). Not a landmark matter, but found while looking at them. -
The laser scan for more sites: Saxony's for Meißen, Grimma and Leipzig, NRW's "3D-Messdaten" for Unna (four 1 km LAZ merged per tile, intensities scaled 1/16) — sheds, extra crowns and hedges there as in Dresden (Meißen 475 small structures and 6 449 extra trees; Grimma 960 and 7 964; Leipzig 1 303 and 27 893; Unna 1 596 and 2 385). The scan could later measure a column's width too (open item below).
What was rejected
- Additions from the surface model alone — cranes (Leipzig's centre, 50–95 m), pylons, crowns. OSM names, the DOM measures (🗃️ ledger).
- Window grids, panes, textures for glass and metal — the standing veto; a cool tint and a sheen only (🗃️ ledger).
- Invented or imported geometry — a stock spire, a hand-modelled landmark, a warehouse glTF: not repeatable, a licence each, off-style (ADR 0036).
- An absolute sitelink threshold — it left Unna with one landmark; the ranking is per tile.
- A fixed count per tile (the first 12) — on a dense tile it cut the second rank; replaced by a floor relative to the tile's top (2026-10-01).
Open items
- GPU plates not judged: every look here (columns, mapped colours, the glass and metal response, the relief) was checked headless only.
- Relief thresholds (3 m above the highest roof, 60 m², 2 %, the smoothing: half the cell, σ = 1 cell) are tuned on synthetic tests and a handful of real cases (the Elbphilharmonie's crests, the Heinrich-Hertz-Turm's top, truncated spires, Unna's hospital — see Findings); the Congress Center's curved roof and a domed church are the next checks.
- Wikidata coverage varies: sitelinks favour the internationally
known; an item without an OSM
wikidata=tag and not a building matches only by point. A small town's list may name a museum over its church. - Column width is a floor and a clamp, not a measurement; the laser point cloud (LSC) would resolve a shaft.
Dense tiles: at most 12 landmarks a tile, so on an old town's tile the most notable win — Dresden's Congress Center is not on the Altstadt tile's list (Frauenkirche, Zwinger, Semperoper, …) and so has neither the flag nor Wikidata's material; its OSM tags still colour it.Resolved 2026-10-01 (plan 051, ADR 0012): a tile keeps every landmark with ≥ 8 % of its most notable one's sitelinks (NOTABLE_SHARE, never under 2), 40 a tile only as a safety cap — a floor relative to the tile, not a count; re-baked on every site. The roof relief stays with the tile's 12 most notable (RELIEF_LANDMARKS): further down the list the buildings are ordinary, and over a villa among old trees the crowns would read as its roof's form. The Congress Center itself is still not on the list: it is not in Wikidata's answer for the tile at all (the query keeps a box's 80 most linked items; the Altstadt tile has more), so the floor cannot reach it — its OSM tags colour it.- The Elbphilharmonie's brick base (the Kaispeicher A, ≈ 37 m) is in
no open source we read: LoD2 is one block, OSM maps the Elbphilharmonie
as one outline without
building:part, and Wikidata names steel, glass and concrete only — so the whole building wears one look. The data to add is an OSMbuilding:partfor the warehouse withheight≈ 37 andbuilding:material=brick. Vertical zoning from OSM parts (a part's height band wearing its own material) is a possible later step, but thin today: 11 parts with both a height and a material in our Hamburg tiles, 1 in Munich. NRW's laser-scan intensity scale is unverifiedResolved 2026-10-01 (plan 051): every scan is normalised against its own ground's median intensity (NRW's factors 0.025–0.027), and NRW's classes are mapped by a table of their own — its crown tops are class 1, which the rasters had dropped.
