#!/usr/bin/env python3 """SpatialGen-Testset layouts -> the unified room-layout schema. One scene -> one record with a single room. The only input is each scene's ``room_layout.json`` plus the caption JSONL; the 4,800 rendered images and the 48 ``cameras.json`` are not downloaded, because the schema has no field for a view or a camera pose. Why each value is what it is -- which published export it copies, and which choices are ours -- is in the README. conda run -n base python kits/convert_spatialgen.py [--limit N] """ from __future__ import annotations import argparse, collections, json, math from pathlib import Path DATASET = "SpatialGen" SPLIT = "test" LICENSE_TAG = "CC BY-NC 4.0 (attribution, non-commercial)" # floor[] and ceil[] are millimetres, bboxes[] are metres -- the two units live # in one file. The factor is the source's own: layout_utils.py sets SCALE=0.001 # before touching floor or ceil, and never scales a box. MILLIMETRE = 0.001 # PROJECT DECISION (approved 2026-09-09): three labels are renamed, written as # "new(source)" so the source string stays readable in the value itself. # # bboxes[].class comes from an ADE20K label set: 60 of the 62 labels in # layout_utils.py carry the exact RGB of an ADE20K-150 class, byte for byte # (only "nightstand" and "other" are the authors' own). ADE20K is a general # scene-parsing vocabulary with no class for a recessed or track-mounted light, # so 372 internal ids collapse onto those 62 labels and the indoor fixtures land # on outdoor words. Checked by projecting the eight corners of each box into # the official renders with the released camera parameters: "traffic light" is # ceiling downlights and track spots (261 boxes, 26.7%), "screen door, screen" # is ordinary interior doors (82), "street lamp" is table and floor lamps (6). CATEGORY_RENAMES = { "traffic light": "ceiling lamp", "screen door, screen": "door", "street lamp": "lamp", } # "chandelier" (ring pendants) and "other" (a genuine mixed bag: vents, chair # groups, plants) were checked the same way and are right, so they stay. So # does "crt screen" -- its one instance is a slatted wall panel that no label in # the vocabulary describes, and inventing one for a single box helps nobody. # Openings are matched on the SOURCE string, before the rename. OPENING_LABELS = {"screen door, screen": "doors", "window": "windows"} # PROJECT DECISION (approved 2026-09-09): room_type read from the caption. # # room_layout.json has no room-type field and room_uid is just "scene_00000", # so the caption is the only place the room is named. Every one of the 48 # captions names it in the first sentence, but a regex tuned to 48 samples is a # lookup table that hides its judgement, so the table is written out instead. # The caption itself is published verbatim in `description`, so every entry here # can be checked against the record that carries it. Values follow IL3D's # vocabulary, the one this project's 3D-FRONT export already normalises to. # # The type is whatever the first sentence calls the room, never the furniture it # goes on to list: scene_00019 opens with "a modern, open-concept living area" # and only later mentions a dining table and a kitchen area, so it is a living # room. A first sentence that declares a blend of living and dining gets IL3D's # compound value: the renders for scene_00012, scene_00044 and scene_00046 all # show a dining table and a seating area in the same space, so the compound is # what the room is, not a wording artefact. ROOM_TYPES = { "scene_00000": "living room", "scene_00001": "bedroom", "scene_00002": "bedroom", "scene_00003": "bathroom", "scene_00004": "bedroom", "scene_00005": "kitchen", "scene_00006": "living room", "scene_00007": "kitchen", "scene_00008": "bedroom", "scene_00009": "living room", "scene_00010": "bedroom", "scene_00011": "bedroom", "scene_00012": "living dining room", "scene_00013": "bedroom", "scene_00014": "bedroom", "scene_00015": "office", "scene_00016": "bathroom", "scene_00017": "living room", "scene_00018": "bedroom", "scene_00019": "living room", "scene_00020": "bedroom", "scene_00021": "bedroom", "scene_00022": "kitchen", "scene_00023": "bedroom", "scene_00024": "office", "scene_00025": "kitchen", "scene_00026": "kitchen", "scene_00027": "living room", "scene_00029": "bedroom", "scene_00030": "bedroom", "scene_00031": "bedroom", "scene_00032": "bedroom", "scene_00033": "bedroom", "scene_00034": "kitchen", "scene_00035": "bedroom", "scene_00036": "bathroom", "scene_00037": "bedroom", "scene_00038": "bedroom", "scene_00039": "bedroom", "scene_00040": "kitchen", "scene_00041": "kitchen", "scene_00042": "kitchen", "scene_00043": "living room", "scene_00044": "living dining room", "scene_00045": "bedroom", "scene_00046": "living dining room", "scene_00047": "bedroom", "scene_00048": "bathroom", } # scene_00028 does not exist: the release numbers 00000-00048 and skips it. # room_type_source per scene. 41 captions name the room with the vocabulary # word itself in their first sentence, which is exactly what this collection # already calls "description_explicit_keyword" (M3DLayout writes it for the # Infinigen and Matterport3D room types it reads out of a description). # # PROJECT DECISION (approved 2026-09-09): the seven below say what the room is # without using that word -- "living space" / "living area" for a living room, # "hotel suite or guest room" for a bedroom, and three first sentences that # declare a blend of dining and living the vocabulary has one compound word for. # The root README's room_type_source table lists "inferred_from_description" for # exactly this -- a room type read out of a description rather than quoted from # it. The published data does not carry that value anywhere (the export the # README attributes it to writes "description_explicit_keyword"), so it is taken # from the documented enum, not from a sibling export. The alternative was the # collection's "unavailable" plus room_type "unknown", which throws away seven # room types the caption does state. ROOM_TYPE_INFERRED = frozenset({ "scene_00006", # "a modern, open-concept living space" "scene_00009", # "a modern living space featuring a cozy seating area" "scene_00011", # "a modern, comfortable hotel suite or guest room" "scene_00012", # "blends the kitchen, dining area, and living room" "scene_00019", # "a modern, open-concept living area" "scene_00044", # "blends a dining area with a living room" "scene_00046", # "a modern dining and living area" }) EXPLICIT_KEYWORD = "description_explicit_keyword" INFERRED = "inferred_from_description" COORDINATE_SYSTEM = ( "meter; right-handed; Z-up; scene-local origin; source axes unchanged; " "furniture_position=box center; furniture_size=instance-scaled local XYZ " "full extents; furniture_rotation=yaw about +Z in degrees, x/y=0 when " "upright and null when tilted, z=null when yaw is undefined; " "source_fields.floor/ceil retain source millimeters" ) # -------------------------------------------------------------------------- # Geometry helpers (standard library only) # -------------------------------------------------------------------------- def matrix_of(flat): """The 16 numbers of `transform` as 4 rows of 4.""" return [[float(flat[r * 4 + c]) for c in range(4)] for r in range(4)] def column_norms(matrix): return [math.sqrt(sum(matrix[r][c] ** 2 for r in range(3))) for c in range(3)] def determinant(m): return (m[0][0] * (m[1][1] * m[2][2] - m[1][2] * m[2][1]) - m[0][1] * (m[1][0] * m[2][2] - m[1][2] * m[2][0]) + m[0][2] * (m[1][0] * m[2][1] - m[1][1] * m[2][0])) def yaw_and_tilt(matrix): """Yaw about the output +Z and a tilt flag. IL3D's rule, and this project's 3D-FRONT converter uses the same function. It reads column 0 for the yaw and column 2 for the tilt, and never reads column 1 -- which is why the 36.8% of source frames whose determinant is -1 need no repair. A negative determinant means the designer mirrored the instance; a box is symmetric about each of its own axes, so the mirror moves none of its eight corners and only flips the sign of an axis this rule does not use. The unmodified transform is kept in source_fields either way. """ horizontal = math.hypot(matrix[0][0], matrix[1][0]) yaw = None if horizontal > 1e-8: yaw = math.degrees(math.atan2(matrix[1][0], matrix[0][0])) yaw = ((yaw + 180.0) % 360.0) - 180.0 yaw = round(yaw, 6) up = [matrix[i][2] for i in range(3)] has_tilt = math.sqrt(up[0] ** 2 + up[1] ** 2 + (up[2] - 1.0) ** 2) > 1e-4 return yaw, has_tilt def chain_loops(segments, tol=1e-6): """Wall segments -> closed loops of XY points. The source lists the segments in no particular order, so they are chained end to end. Each loop is returned without the repeated closing vertex, or None in place of a loop that does not close. """ remaining = list(segments) used = [False] * len(remaining) def same(p, q): return abs(p[0] - q[0]) <= tol and abs(p[1] - q[1]) <= tol loops = [] for seed in range(len(remaining)): if used[seed]: continue used[seed] = True points = [remaining[seed][0], remaining[seed][1]] growing = True while growing: growing = False for i, (a, b) in enumerate(remaining): if used[i]: continue if same(points[-1], a): points.append(b) elif same(points[-1], b): points.append(a) elif same(points[0], b): points.insert(0, a) elif same(points[0], a): points.insert(0, b) else: continue used[i] = True growing = True loops.append(points[:-1] if same(points[0], points[-1]) else None) return loops def signed_area(points): """Shoelace. Positive when the ring is counter-clockwise.""" total = 0.0 for i, (x, y) in enumerate(points): nx, ny = points[(i + 1) % len(points)] total += x * ny - nx * y return total / 2.0 def canonical_ring(points): """Counter-clockwise, starting at the lexicographically smallest vertex. A ring has as many encodings as it has vertices, times two for the winding. Pinning one of them is what lets a value be compared between runs; it is a choice of encoding, not a change of shape. Matches the 3D-FRONT converter, whose hull also starts at the smallest vertex. """ if signed_area(points) < 0: points = points[::-1] start = min(range(len(points)), key=lambda i: points[i]) return points[start:] + points[:start] # -------------------------------------------------------------------------- # Records # -------------------------------------------------------------------------- def category_of(raw): """Renamed labels keep the source string in brackets; the rest are the source string with its whitespace trimmed -- the window label carries a trailing space, an artefact of shortening ADE20K's "windowpane;window".""" new = CATEGORY_RENAMES.get(raw) return f"{new}({raw})" if new else raw.strip() def furniture_record(box, index, scene_id, counters): matrix = matrix_of(box["transform"]) norms = column_norms(matrix) frame = [[matrix[r][c] / norms[c] for c in range(3)] for r in range(3)] yaw, has_tilt = yaw_and_tilt(frame) mirrored = determinant(frame) < 0 # Rounded once, so the counter below and the published record agree on # what "flat" means; counting the unrounded value disagrees on one box. size = [round(float(box["size"][i]) * norms[i], 6) for i in range(3)] counters["furniture_mirrored_frame"] += mirrored counters["furniture_tilted"] += has_tilt counters["furniture_flat"] += min(size) < 0.02 counters["furniture_renamed"] += box["class"] in CATEGORY_RENAMES return { "furniture_category": category_of(box["class"]), # No source carries an object id: bboxes[] entries hold class, transform # and size and nothing else, so the index in that list is the only # handle the source gives. "furniture_instance_id": f"{scene_id}::obj_{index}", "furniture_position": dict(zip("xyz", (round(matrix[i][3], 6) for i in range(3)))), # x and y are 0.0 for an upright box and null for a tilted one, rather # than an invented angle -- IL3D aggregate_scenes.py:245. "furniture_rotation": {"x": None if has_tilt else 0.0, "y": None if has_tilt else 0.0, "z": yaw}, "furniture_size": dict(zip(("width", "length", "height"), size)), # Every source key, unmodified, plus the two derived numbers needed to # read the published values back: the scale that separates size from # furniture_size, and the handedness that furniture_rotation drops. "source_fields": { "class": box["class"], "transform": box["transform"], "size": box["size"], "index": index, "transform_scale": [round(v, 9) for v in norms], "transform_mirrored": mirrored, }, } def convert_scene(scene_id, layout, caption, counters): prefixed = f"spatialgen_{scene_id}" boxes = layout["bboxes"] counters["boxes_total"] += len(boxes) furniture, openings = [], collections.defaultdict(list) for index, box in enumerate(boxes): furniture.append(furniture_record(box, index, prefixed, counters)) field = OPENING_LABELS.get(box["class"].strip()) if field: openings[field].append({"furniture_index": index}) for field in ("doors", "windows"): counters[f"empty_{field}"] += not openings[field] segments = [((round(s["start"]["x"] * MILLIMETRE, 9), round(s["start"]["y"] * MILLIMETRE, 9)), (round(s["end"]["x"] * MILLIMETRE, 9), round(s["end"]["y"] * MILLIMETRE, 9))) for s in layout["floor"]] loops = chain_loops(segments) if len(loops) == 1 and loops[0] is not None and len(loops[0]) >= 3: # Rounded first, then canonicalised, so the published ring really does # start at its own smallest vertex: ordering the full-precision points # and rounding afterwards can put two vertices that tie at six decimals # the other way round. The area follows the published ring for the same # reason -- recomputing it from room_boundary has to give room_area_m2. ring = canonical_ring([(round(x, 6), round(y, 6)) for x, y in loops[0]]) boundary = [[x, y] for x, y in ring] area = round(abs(signed_area(ring)), 6) boundary_type = "polygon" else: # Not reached on the 48 released scenes; every one closes into a single # loop. A future release that does not gets null, not a hull. boundary, area, boundary_type = None, None, None counters["scenes_without_single_loop"] += 1 # The source's own room height: layout_utils.py averages the two rings and # subtracts. Per-segment z is kept in source_fields, so the spread stays # visible even though only the mean is published. floor_z = [s[end]["z"] for s in layout["floor"] for end in ("start", "end")] ceil_z = [s[end]["z"] for s in layout["ceil"] for end in ("start", "end")] height = round((sum(ceil_z) / len(ceil_z) - sum(floor_z) / len(floor_z)) * MILLIMETRE, 6) room = { "room_id": f"{prefixed}::room_0", "room_index": 0, "room_type": ROOM_TYPES.get(scene_id), "room_type_source": (INFERRED if scene_id in ROOM_TYPE_INFERRED else EXPLICIT_KEYWORD), "room_boundary": boundary, "room_boundary_type": boundary_type, "room_area_m2": area, "room_height": height, "coordinate_system": COORDINATE_SYSTEM, "doors": openings["doors"], "windows": openings["windows"], "furniture_quantity": len(furniture), "furniture": furniture, "source_fields": { "room_uid": layout.get("room_uid"), # Consumed into room_boundary and room_height, but lossily: the # segment pairing and the per-segment z do not survive either. "floor": layout["floor"], "ceil": layout["ceil"], }, } counters["room_type_null"] += room["room_type"] is None counters["room_type_inferred"] += scene_id in ROOM_TYPE_INFERRED return { "scene_id": prefixed, "sample_id": prefixed, "dataset": DATASET, "source_dataset": DATASET, "source_subset": SPLIT, "room_count": 1, "total_area": area, "building_style": None, "description": caption, # Every published export writes null here, and the one that does not # (M3DLayout) writes the parsed source description object, not a path. # The caption itself is already published verbatim in `description`. "description_source": None, "provenance": { "source_file": f"data/{scene_id}/room_layout.json", "split": SPLIT, "license_tag": LICENSE_TAG, }, "rooms": [room], } def main(): here = Path(__file__).resolve().parent.parent ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) ap.add_argument("--data-root", type=Path, default=here / "data") ap.add_argument("--out", type=Path, default=here / "spatialgen.jsonl") ap.add_argument("--limit", type=int, help="stop after N scenes -- use before any full run") args = ap.parse_args() captions = {} with open(args.data_root / "test_split_caption.jsonl", encoding="utf-8") as handle: for line in handle: if line.strip(): captions.update(json.loads(line)) split_file = args.data_root / "test_split.txt" scenes = [line.strip() for line in split_file.read_text().splitlines() if line.strip()] if args.limit: scenes = scenes[:args.limit] counters = collections.Counter() written = 0 tmp = args.out.with_suffix(args.out.suffix + ".part") with open(tmp, "w", encoding="utf-8") as out: for scene_id in scenes: path = args.data_root / scene_id / "room_layout.json" if not path.exists(): counters["scenes_missing_layout"] += 1 continue layout = json.loads(path.read_text(encoding="utf-8")) caption = captions.get(scene_id) counters["captions_missing"] += caption is None record = convert_scene(scene_id, layout, caption, counters) out.write(json.dumps(record, ensure_ascii=False) + "\n") written += 1 tmp.replace(args.out) counters["scenes_written"] = written print(json.dumps(dict(sorted(counters.items())), indent=2)) if __name__ == "__main__": main()