ResPlan-FP

A floorplan-vectorization benchmark of 16,998 residential plans rendered as clean 256 px raster images with wall-first vector ground truth (walls as centerline segments with thickness, openings on walls, rooms as wall-cycle polygons with labels). It accompanies the paper

When Should a Network Emit Geometry, and When Should It Detect It? Readout, Reconciliation, and Representation in Floorplan Vectorization — He Zhang, arXiv:2608.25608

Code (converter, render protocol, split freezing, evaluator, baseline scoring): https://github.com/Cyprinus12138/fpvec-lab

Files

file	content
train.jsonl / val.jsonl / test.jsonl	one row per plan: id, gt (wall-first JSON, 1024 longer-edge frame), roles, n_walls, n_rooms, n_openings, watertight, r_val, image, style
images_{train,val,test}.tar	images/resplan_<id>.png, 256 px clean_band_v1 renders
splits.json	seed, counts (14,998 / 1,000 / 1,000), per-split id fingerprints and the full test-id list
stats.json	conversion statistics
baselines/*.preds.jsonl	predictions of the paper's baseline tracks on the test split ({"id", "pred"} per line)
baselines/llm_subset_ids.json	the 200-plan subset used for the LLM track

Split protocol

Accepted plan ids sorted as strings, shuffled with random.Random(42); first 1,000 = test, next 1,000 = val, rest = train. Fingerprint of a split = sha256(",".join(sorted(ids)))[:16]:

split	n	fingerprint
test	1000	669656eabb515960
val	1000	efb5e82b808acec4
train	14998	5af77996b313bba2

python -m resplan_fp.build in the code repo rebuilds the benchmark from ResPlan.pkl and aborts if the fingerprints differ.

Render protocol (clean_band_v1)

White 256×256 canvas; each wall drawn as a solid dark band of one per-plan thickness max(5, round(median wall thickness × 256/1024)) px; every opening cut out of the band as a white gap; doors get a quarter-circle swing arc and leaf, windows a thin double line across the gap. No text, no fixtures, no noise. The renderer in the code repo reproduces the shipped PNGs byte for byte.

Baseline tracks (wall F1 on the test split, all 1,000 plans, missing predictions score 0)

track	file	wall F1 @0.05	@0.015
zero-shot, wall-first sequence decode	zeroshot_seq_decode	0.640	0.505
zero-shot, graph readout, thresholds frozen on CubiCasa5K	zeroshot_readout_frozen	0.621	0.553
zero-shot, graph readout, thresholds calibrated on ResPlan-FP val	zeroshot_readout_calibrated	0.677	—
zero-shot, room-centric (Raster2Seq continued)	zeroshot_roomcentric	0.383	—
zero-shot, fusion (room-centric + readout donor, readout fallback)	zeroshot_fusion	0.688	—
finetuned, wall-first sequence decode	finetuned_seq_decode	0.968	0.957
finetuned, graph readout (val-calibrated)	finetuned_readout	0.915	—
finetuned, room-centric (Raster2Seq finetuned)	finetuned_roomcentric	0.928	0.902
LLM zero-shot, Gemini 3.1 Pro, 200-plan subset	llm_gemini	0.811	0.472

Score any file with python -m baselines.score --bench <this dir> --preds baselines/<file>.preds.jsonl. The LLM file keeps the model's declared rooms; the paper's room / r_val values for that row were computed with the model's rooms discarded (rooms read as faces of the predicted wall graph), which is --drop-pred-rooms in the scorer.

Known limitations of the conversion

• Open-plan kitchens that share a wall-graph face with the living room are labelled living (max-overlap labelling); kitchen polygons exist in the source for future sub-face labels.
• Balconies are absent: ResPlan railings are not in the wall layer, so they form no face.
• 2 of 17,000 source plans are rejected (more than 80 walls or more than 512 sequence tokens). Watertightness 0.995.

License and provenance

Derived from ResPlan (Abouagour et al.; data CC BY 4.0, code MIT). ResPlan-FP is released under CC BY 4.0.

Citation

@article{zhang2026emitdetect,
  title   = {When Should a Network Emit Geometry, and When Should It Detect It?
             Readout, Reconciliation, and Representation in Floorplan Vectorization},
  author  = {Zhang, He},
  journal = {arXiv preprint arXiv:2608.25608},
  year    = {2026}
}