FloorPlanCAD primitive graphs: Visionaries pre-training subset

Graph-structured training data for a GAT-CADNet-style panoptic symbol spotting model (CV5570, Team Visionaries). Each sample is one tile of a FloorPlanCAD drawing: nodes are vector primitives (line, arc, circle, ellipse, curve), edges link primitives within 300 mm plus collinear neighbours, capped at 30 per node and 4096 nodes per tile.

Source and license. Derived from FloorPlanCAD (Fan et al., ICCV 2021, https://floorplancad.github.io/). The annotations are CC BY-NC 4.0, and the dataset authors state they do not own the copyright of the drawings. This derivative is therefore non-commercial only and is kept private by default. No raw SVGs are included unless INCLUDE_RAW_SVGS was enabled.

Splits

split	drawings	tiles	nodes	edges	edge pos_weight
train	2482	2887	2,477,634	29,188,839	1.9491
val	298	341	273,545	3,248,997	1.6439
test	5502	5856	3,966,895	43,072,174	2.0321

train is a stratified 2500-drawing subset of the FloorPlanCAD training pool (645 chosen greedily to cover rare classes, the rest size-stratified random). val is drawn from held-out projects. test comes from the official test split. The file lists are in splits/.

Taxonomy (harmonised from the 35 FloorPlanCAD classes)

id	class	kind	FloorPlanCAD source classes
0	wall	stuff	wall
1	curtain_wall	stuff	curtain wall
2	railing	stuff	railing
3	door_single	thing	single door
4	door_double	thing	double door
5	door_sliding	thing	sliding door
6	door_other	thing	folding door, revolving door, rolling door
7	window	thing	window, bay window, blind window
8	opening	thing	opening symbol
9	stairs	thing	stairs
10	elevator	thing	elevator
11	other	thing	-

Everything not listed, including unlabelled primitives, maps to other. Instance ids exist only for thing classes.

Files

• graphs/<split>/<split>-NNNNN.npz: shards of concatenated tiles. Slice a tile with node_ptr and edge_ptr.
  • node arrays: x [N,13] float32, pts [N,4,2] (4 sampled points, tile-local mm), ptype, y (class id), inst (-1 = none), raw (original FloorPlanCAD id, 0-based, 35 = unlabelled)
  • edge arrays: edge_index [2,E] with i<j (undirected, stored once), edge_attr [E,8], edge_y (same instance), edge_mask (edge counts in the instance loss)
• index_<split>.parquet: one row per tile (drawing, window, shard, position, sizes, per-class node counts).
• stats.json: class counts and weights, edge pos_weight, validation problems. config.json: every parameter used.
• fpgraph.py: the parser and graph builder. Use the same file for fine-tuning and inference.

Node features: is_line, is_arc, is_circle, is_ellipse, is_curve, log_length_m, cos2theta, sin2theta, sweep_frac, is_closed, cx_tile, cy_tile, log_radius_m. Edge features: min_dist_rel, centre_dist_m, cos_dtheta, abs_sin_dtheta, log_len_ratio, touch, collinear, perp_offset_m. Stroke colour is deliberately not a feature, because FloorPlanCAD colours encode the class.

from fpgraph import load_shard, get_tile, to_pyg
tile = get_tile(load_shard("graphs/train/train-00000.npz"), 0)
data = to_pyg(tile)