4.0 KiB
Buildings not yet voxelized
Status of the 34 catalogued drawings against the buildings list in
data/voxel_heights.lua. Coverage is 31 of 34 drawings, 144 of 147
placements. Three are left, one placement each, and none of them is a
matter of sitting down and reading the drawing - each defeats a different
assumption the band-table pipeline is built on.
(Counting note: the catalogue's per-building tables list one row per DOOR, so B19, B22 and B23 each appear twice for a single placement. 150 rows, 147 placements. This page counts placements.)
| id | tileset | cells | px | used | what defeats it |
|---|---|---|---|---|---|
| B19 | PLATEAU |
20 x 5 | 320x80 | 1x | two structures in one drawing |
| B30 | OVERWORLD |
6 x 4 | 96x64 | 1x | truncated by the map edge; no roof drawn |
| B32 | SHIP_PORT |
8 x 3 | 128x48 | 1x | not a building |
The silhouette test
A drawing the pipeline can read floods to one connected piece. That is the most useful single number here, more than fill percentage: a drawing can be 95% filled and still be shredded.
| id | fill | pieces | largest piece | note |
|---|---|---|---|---|
| shipped, for reference | 82-95% | 1 | 100% | |
| B19 | 76% | 1 | 100% | silhouette is fine; the problem is elsewhere |
| B32 | 95% | 29 | 98% | mostly intact, fragments are the ship's fittings |
| B30 | 37% | 126 | 25% | unbounded - see below |
B19 - Indigo Plateau
INDIGO_PLATEAU (0,1), doors at (9,5) and (10,5) into
INDIGO_PLATEAU_LOBBY.
Its silhouette is clean, and B23 - the Victory Road entrance, the same tileset and the same kind of rock face - was voxelized without trouble. The difference is that B19 is two structures in one drawing. Rows 0..47 span the full 320px: that is the plateau wall. Rows 48..79 span only x=96..223: that is the lobby building standing in front of it, 8 cells wide.
The band table describes one roof over one facade. Give B19 a single
roofRows and the roof slab covers all 320 columns while walls exist only
under the middle 128, so the two ends come out as a slab floating over
nothing. Nothing in the band table can say "and the wall stops here".
It needs either a way to express two stacked footprints in one template, or splitting into two drawings - which means changing the extraction, not the profile.
B30 - the Pokemon Tower
LAVENDER_TOWN (12,0). A tall face of windows over brick.
Two problems, and the second is the one that stops it.
Its silhouette is unbounded. The drawing has no black outline anywhere
on its boundary: row 0, row H-1 and both side columns are entirely light,
so the flood enters at 318 separate border seeds and eats everything up to
the window frames - 126 fragments, largest 25%. seal does not rescue it
the way it did Route 10's block: sealing the south side alone changes
nothing (37% either way), and sealing all four asserts the whole box is
building.
It cannot be sealed, because the box is not all building. The catalogue's own silhouette record says 1552 of the 6144 px are ground - a quarter of the box - including interior tile columns. Seal all four sides and that quarter becomes wall.
And it has no roof. The tower sits at y=0, the top row of the map, so
the drawing is cut off by the map boundary: rows 0..59 are windows and
brick all the way up, with no roof band. A band table with a small
roofRows caps it with its own top rows, which renders plausibly, but the
cap is a reading the drawing does not support.
B32 - the S.S. Anne
VERMILION_DOCK (10,3). A ship in three-quarter perspective, with a hull,
deck, funnels and gangway.
It is the only asymmetric drawing in the catalogue (top[x] does not
mirror), which by itself fails the reference tool's symmetry assert. Its
silhouette is largely fine, but the band table does not describe it at any
setting: there is no roof-from-above band, no face-on facade, and no taper
that means elevation. Voxelizing the S.S. Anne means a different archetype,
not a band table.