Fire3DSingleImageData / build_gt_from_3dfront.py
TianhangCheng7's picture
gather gt from full dataset (part 3)
46f479f verified
Raw
History Blame Contribute Delete
18.8 kB
#!/usr/bin/env python
"""Build per-view GT (instance mask + object meshes) for the Fire3D single_image
scenes by aligning the source 3D-FRONT room to each annotation and ray-casting it
into the annotation camera.
Pipeline per scene (see the discussion in the task):
1. shortlist candidate 3D-FRONT rooms whose furniture-model UUID set covers the
annotation's objects (index built from the ``*_full.glb`` scene graphs);
2. estimate the glb->annotation-world similarity transform with a RANSAC over
per-model candidate matches (robust to duplicate models / extra room objects);
3. for the top-K candidates, ray-cast the placed room objects through the
annotation intrinsics and keep the room whose rendered depth best agrees with
the dataset metric depth;
4. write outputs: uint16 instance mask (full-res, aligned to rgb), id->object
json, per-object meshes (PLY, OpenCV camera frame = the sceneobjgt frame),
merged scene-objects mesh, and an rgb overlay for eyeballing the match.
Objects only: walls / floor / ceiling are intentionally skipped.
"""
from __future__ import annotations
import argparse
import glob
import json
import os
import re
import warnings
from itertools import combinations, product
from typing import Any
import numpy as np
import trimesh
from PIL import Image
warnings.filterwarnings("ignore")
UUID_RE = re.compile(r"([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})")
SCENE_ROOT = "/mnt/task_runtime/3d-front/3D-FRONT-SCENE"
SI_ROOT = "/mnt/task_runtime/data/single_image"
CV2GL = np.diag([1.0, -1.0, -1.0]) # OpenCV cam <-> OpenGL cam
CANONICAL_R = np.array([[1.0, 0, 0], [0, 0, -1.0], [0, 1.0, 0]]) # glb y-up -> world z-up
PALETTE = np.array([
[230, 25, 75], [60, 180, 75], [255, 225, 25], [0, 130, 200], [245, 130, 48],
[145, 30, 180], [70, 240, 240], [240, 50, 230], [210, 245, 60], [250, 190, 212],
[0, 128, 128], [220, 190, 255], [170, 110, 40], [255, 250, 200], [128, 0, 0],
[170, 255, 195], [128, 128, 0], [255, 215, 180], [0, 0, 128], [128, 128, 128],
], dtype=np.uint8)
def umeyama(src: np.ndarray, dst: np.ndarray, fixed_r: np.ndarray | None = None
) -> tuple[float, np.ndarray, np.ndarray]:
"""Fit a similarity transform ``dst ~= s * R @ src + t`` (Umeyama, 1991).
If ``fixed_r`` is given the rotation is held fixed and only scale/translation
are estimated (used when only two correspondences are available).
"""
mu_s, mu_d = src.mean(0), dst.mean(0)
s_c, d_c = src - mu_s, dst - mu_d
if fixed_r is None:
cov = d_c.T @ s_c / len(src)
u, d, vt = np.linalg.svd(cov)
rot = u @ vt
if np.linalg.det(rot) < 0:
u[:, -1] *= -1
rot = u @ vt
scale = float(np.trace(np.diag(d)) / ((s_c ** 2).sum() / len(src)))
else:
rot = fixed_r
scale = float((d_c * (s_c @ rot.T)).sum() / (s_c ** 2).sum())
trans = mu_d - scale * rot @ mu_s
return scale, rot, trans
def load_room_objects(scene_uuid: str, room: str) -> dict[str, tuple[trimesh.Trimesh, str]]:
"""Load furniture nodes of ``<room>_full.glb`` as ``node_name -> (mesh, model_uuid)``.
Architecture / unnamed geometry (walls, floor, ceiling) carry no model UUID in
the node name and are skipped, so only objects are returned.
"""
full = os.path.join(SCENE_ROOT, scene_uuid, f"{room}_full.glb")
if not os.path.exists(full):
return {}
scene = trimesh.load(full)
nodes: dict[str, tuple[trimesh.Trimesh, str]] = {}
for name in scene.graph.nodes_geometry:
m = UUID_RE.search(name)
if not m:
continue
transform, geom_name = scene.graph[name]
geom = scene.geometry[geom_name].copy()
geom.apply_transform(transform)
nodes[name] = (geom, m.group(1))
return nodes
def fit_similarity(ann: dict[str, Any],
nodes: dict[str, tuple[trimesh.Trimesh, str]],
thresh: float = 0.15) -> dict[str, Any] | None:
"""RANSAC estimate of the glb->annotation-world similarity transform.
Each annotation object may match several same-model room nodes; we sample
triples of (object, candidate-node) hypotheses, fit a similarity, and score it
by the number of annotation objects whose nearest same-model node lands within
``thresh`` metres. Returns the transform plus the inlier node->object assignment.
"""
ann_obj = [(o["model_file_name"][0], np.array(o["bbox3d_world_center"], float), oid)
for oid, o in enumerate(ann["obj_dict"].values())]
node_by_uuid: dict[str, list[tuple[str, np.ndarray]]] = {}
for name, (geom, mid) in nodes.items():
node_by_uuid.setdefault(mid, []).append((name, geom.bounds.mean(0)))
cand = [[(nm, c) for nm, c in node_by_uuid.get(mid, [])] for mid, _, _ in ann_obj]
have = [i for i, c in enumerate(cand) if c]
if len(have) < 2:
return None
def score(scale: float, rot: np.ndarray, trans: np.ndarray
) -> tuple[list[float], dict[str, int]]:
res: list[float] = []
assign: dict[str, int] = {}
for i in have:
dst = ann_obj[i][1]
nm, c = min(cand[i], key=lambda nc: np.linalg.norm(scale * (rot @ nc[1]) + trans - dst))
d = float(np.linalg.norm(scale * (rot @ c) + trans - dst))
if d < thresh:
res.append(d)
assign[nm] = ann_obj[i][2]
return res, assign
best: tuple[tuple[int, float], float, np.ndarray, np.ndarray, list[float], dict[str, int]] | None = None
if len(have) >= 3:
trip = sorted(have, key=lambda i: len(cand[i]))[:min(len(have), 6)]
for combo in combinations(trip, 3):
for picks in product(*[cand[i] for i in combo]):
src = np.array([p[1] for p in picks])
dst = np.array([ann_obj[i][1] for i in combo])
if not (np.isfinite(src).all() and np.isfinite(dst).all()):
continue
try:
scale, rot, trans = umeyama(src, dst)
except np.linalg.LinAlgError:
continue
res, assign = score(scale, rot, trans)
key = (len(res), -float(np.sum(res)) if res else 0.0)
if best is None or key > best[0]:
best = (key, scale, rot, trans, res, assign)
if best is None: # two-object scene: fix the canonical rotation
src = np.array([cand[i][0][1] for i in have])
dst = np.array([ann_obj[i][1] for i in have])
gm = np.isfinite(src).all(1) & np.isfinite(dst).all(1)
if gm.sum() < 2:
return None
try:
scale, rot, trans = umeyama(src[gm], dst[gm], CANONICAL_R)
except np.linalg.LinAlgError:
return None
res, assign = score(scale, rot, trans)
best = ((len(res), 0.0), scale, rot, trans, res, assign)
_, scale, rot, trans, res, assign = best
if not res:
return None
return dict(scale=scale, rot=rot, trans=trans, n_inliers=len(res),
max_res=float(np.max(res)), med_res=float(np.median(res)), assign=assign)
def camera_from_annotation(ann: dict[str, Any]) -> tuple[np.ndarray, np.ndarray]:
"""Return ``(K, T_cv_from_world)``: intrinsics and world->OpenCV-camera 4x4."""
k = np.asarray(ann["camera_intrinsics"], float)
w2c = np.eye(4)
w2c[:3] = np.asarray(ann["camera_extrinsics"], float) # world -> OpenGL cam
rot, trans = w2c[:3, :3], w2c[:3, 3]
t_world_from_cv = np.eye(4)
t_world_from_cv[:3, :3] = rot.T @ CV2GL
t_world_from_cv[:3, 3] = -rot.T @ trans
return k, np.linalg.inv(t_world_from_cv)
def place_objects(nodes: dict[str, tuple[trimesh.Trimesh, str]], fit: dict[str, Any],
t_cv_from_world: np.ndarray
) -> list[tuple[str, str, trimesh.Trimesh]]:
"""Transform each room object node glb->world (similarity)->OpenCV camera frame.
Returns a deterministic list of ``(node_name, model_uuid, mesh_in_camera_frame)``.
"""
scale, rot, trans = fit["scale"], fit["rot"], fit["trans"]
placed: list[tuple[str, str, trimesh.Trimesh]] = []
for name in sorted(nodes):
geom, mid = nodes[name]
mesh = geom.copy()
mesh.vertices = scale * (rot @ mesh.vertices.T).T + trans # -> world
mesh.apply_transform(t_cv_from_world) # -> camera (cv)
placed.append((name, mid, mesh))
return placed
def raycast_instance_mask(placed: list[tuple[str, str, trimesh.Trimesh]],
k: np.ndarray, hw: tuple[int, int], stride: int = 1
) -> tuple[np.ndarray, np.ndarray]:
"""Ray-cast placed objects through ``K``; return ``(instance_id_map, z_depth)``.
Instance ids are ``1..len(placed)`` (0 = background); the nearest surface wins
per pixel. ``z_depth`` is the camera-frame z of the hit (inf where no hit).
"""
height, width = hw
faces_obj = np.concatenate([np.full(len(m.faces), i + 1, np.int32)
for i, (_, _, m) in enumerate(placed)])
combined = trimesh.util.concatenate([m for _, _, m in placed])
fx, fy, cx, cy = k[0, 0], k[1, 1], k[0, 2], k[1, 2]
us, vs = np.meshgrid(np.arange(0, width, stride), np.arange(0, height, stride))
flat_u, flat_v = us.ravel(), vs.ravel()
dirs = np.stack([(flat_u - cx) / fx, (flat_v - cy) / fy, np.ones_like(flat_u, float)], -1)
dirs /= np.linalg.norm(dirs, axis=1, keepdims=True)
origins = np.zeros_like(dirs)
loc, ray_idx, tri_idx = combined.ray.intersects_location(origins, dirs, multiple_hits=False)
id_flat = np.zeros(flat_u.shape, np.int32)
z_flat = np.full(flat_u.shape, np.inf)
for r, tri, xyz in zip(ray_idx, tri_idx, loc):
if xyz[2] < z_flat[r]:
z_flat[r] = xyz[2]
id_flat[r] = faces_obj[tri]
out_h, out_w = us.shape
return id_flat.reshape(out_h, out_w), z_flat.reshape(out_h, out_w)
def depth_agreement(z_render: np.ndarray, depth: np.ndarray) -> dict[str, float]:
"""Agreement between rendered object z and dataset metric depth on covered pixels."""
valid = np.isfinite(z_render) & np.isfinite(depth) & (depth > 0) & (depth < 100)
if valid.sum() == 0:
return dict(coverage=0.0, med_mm=float("inf"), within5mm=0.0, n=0)
err = np.abs(z_render[valid] - depth[valid])
return dict(coverage=float(np.isfinite(z_render).mean()),
med_mm=float(np.median(err) * 1000), within5mm=float((err < 0.005).mean()),
n=int(valid.sum()))
def shortlist_rooms(ann: dict[str, Any], u2rooms: dict[str, set[str]]) -> list[str]:
"""Rooms whose model set covers all annotation models (else the best-covering ones)."""
uuids = set(o["model_file_name"][0] for o in ann["obj_dict"].values())
sets = [u2rooms.get(u, set()) for u in uuids]
if sets and all(sets):
common = set.intersection(*sets)
if common:
return sorted(common)
from collections import Counter
counter: Counter[str] = Counter()
for s in sets:
counter.update(s)
if not counter:
return []
top = max(counter.values())
return sorted(r for r, n in counter.items() if n == top)
def process_scene(scene_id: str, u2rooms: dict[str, set[str]], out_root: str,
top_k: int, sel_stride: int, min_conf: float = 0.5) -> dict[str, Any]:
"""Match one scene to its 3D-FRONT room and, if confident, write GT into the
scene folder next to the original files, using the ``_<index06>`` naming.
Low-confidence matches (rendered depth disagrees with the metric depth) are
not written; the scene is reported as ``low-confidence`` instead.
"""
scene_dir = os.path.join(out_root, scene_id)
idx6 = f"{int(scene_id):06d}"
ann = json.load(open(glob.glob(f"{scene_dir}/annotation_*.json")[0]))
k, t_cv_from_world = camera_from_annotation(ann)
depth = np.load(glob.glob(f"{scene_dir}/depth_*.npy")[0]).astype(np.float64)
height, width = depth.shape
candidates = shortlist_rooms(ann, u2rooms)
scored: list[tuple[tuple[int, float], str, str, dict[str, Any]]] = []
for room_key in candidates:
scene_uuid, room = room_key.split("|")
nodes = load_room_objects(scene_uuid, room)
if not nodes:
continue
fit = fit_similarity(ann, nodes)
if fit is None:
continue
scored.append(((-fit["n_inliers"], fit["max_res"]), scene_uuid, room, fit))
if not scored:
return dict(scene=scene_id, status="no-match", n_candidates=len(candidates))
scored.sort(key=lambda x: x[0])
# verify the top-K by rendered-depth agreement (decisive for ambiguous scenes)
best = None
for _, scene_uuid, room, fit in scored[:top_k]:
nodes = load_room_objects(scene_uuid, room)
placed = place_objects(nodes, fit, t_cv_from_world)
_, z_low = raycast_instance_mask(placed, k, (height, width), stride=sel_stride)
agree = depth_agreement(z_low, depth[::sel_stride, ::sel_stride])
key = (agree["within5mm"], -agree["med_mm"])
if best is None or key > best[0]:
best = (key, scene_uuid, room, fit, agree)
_, scene_uuid, room, fit, _ = best
# full-res render for the winner
nodes = load_room_objects(scene_uuid, room)
placed = place_objects(nodes, fit, t_cv_from_world)
id_map, z_render = raycast_instance_mask(placed, k, (height, width), stride=1)
agree = depth_agreement(z_render, depth)
# id -> object metadata (computed before writing so we can gate on confidence)
ann_by_oid = list(ann["obj_dict"].values())
id_records: list[dict[str, Any]] = []
for i, (name, mid, _) in enumerate(placed):
inst = i + 1
category = UUID_RE.split(name)[0].strip("_")
oid = fit["assign"].get(name)
label = ann_by_oid[oid]["label"][0] if oid is not None else category
n_pix = int((id_map == inst).sum())
id_records.append(dict(instance_id=inst, node_name=name, model_uuid=mid,
category=category, label=label, annotated=oid is not None,
obj_id=int(ann_by_oid[oid]["obj_id"][0]) if oid is not None else None,
n_pixels=n_pix, visible=n_pix > 0))
confident = agree["within5mm"] >= min_conf
result = dict(scene=scene_id, status="ok" if confident else "low-confidence",
scene_uuid=scene_uuid, room=room, confident=confident,
n_candidates=len(candidates), n_objects=len(placed),
n_annotated=sum(r["annotated"] for r in id_records),
fit=dict(scale=fit["scale"], n_inliers=fit["n_inliers"],
max_res=fit["max_res"], med_res=fit["med_res"]),
depth=agree, instances=id_records)
if not confident:
return result # do not write GT for an unreliable match
# write GT into the scene folder, matching the dataset's _<index06> convention
obj_dir = os.path.join(scene_dir, f"objects_{idx6}")
os.makedirs(obj_dir, exist_ok=True)
Image.fromarray(id_map.astype(np.uint16)).save(os.path.join(scene_dir, f"instance_{idx6}.png"))
merged: list[trimesh.Trimesh] = []
for (name, _, mesh), rec in zip(placed, id_records):
safe = re.sub(r"[^0-9a-zA-Z]+", "_", rec["label"]).strip("_")[:32]
mesh.export(os.path.join(obj_dir, f"{rec['instance_id']:03d}_{safe}.ply"))
merged.append(mesh)
trimesh.util.concatenate(merged).export(os.path.join(scene_dir, f"sceneobjfull_{idx6}.ply"))
save_overlay(scene_dir, id_map, os.path.join(scene_dir, f"instance_overlay_{idx6}.png"))
json.dump({k2: v for k2, v in result.items() if k2 != "instances"} | {
"similarity": dict(scale=fit["scale"], rot=fit["rot"].tolist(), trans=fit["trans"].tolist()),
"instances": id_records}, open(os.path.join(scene_dir, f"instance_{idx6}.json"), "w"), indent=2)
return result
def save_overlay(scene_dir: str, id_map: np.ndarray, path: str) -> None:
"""Blend the coloured instance mask over the rgb image for visual inspection."""
rgb_path = (glob.glob(f"{scene_dir}/rgb_*.jpeg") + glob.glob(f"{scene_dir}/rgb_*.png"))[0]
rgb = np.asarray(Image.open(rgb_path).convert("RGB"))
if rgb.shape[:2] != id_map.shape:
rgb = np.asarray(Image.fromarray(rgb).resize((id_map.shape[1], id_map.shape[0])))
color = np.zeros_like(rgb)
for inst in np.unique(id_map):
if inst == 0:
continue
color[id_map == inst] = PALETTE[(inst - 1) % len(PALETTE)]
fg = id_map > 0
out = rgb.copy()
out[fg] = (0.5 * rgb[fg] + 0.5 * color[fg]).astype(np.uint8)
Image.fromarray(out).save(path)
def main() -> None:
"""CLI entry point: match and export GT for one or all single_image scenes."""
ap = argparse.ArgumentParser(description=__doc__)
ap.add_argument("--scenes", nargs="*", default=None, help="scene ids (default: all)")
ap.add_argument("--index", default="/tmp/uuid_index.json",
help="model-uuid -> [[scene,room],...] index json")
ap.add_argument("--out", default=SI_ROOT,
help="dataset root; GT is written into each <out>/<scene>/ folder")
ap.add_argument("--top-k", type=int, default=8, help="candidates to depth-verify")
ap.add_argument("--sel-stride", type=int, default=4, help="pixel stride for verification")
args = ap.parse_args()
idx = json.load(open(args.index))
u2rooms = {u: set(f"{a}|{b}" for a, b in v) for u, v in idx.items()}
scenes = args.scenes or [os.path.basename(p) for p in sorted(glob.glob(f"{args.out}/[0-9]*"))]
summary = []
for sid in scenes:
try:
res = process_scene(sid, u2rooms, args.out, args.top_k, args.sel_stride)
except Exception as exc: # noqa: BLE001 - keep the batch going
res = dict(scene=sid, status=f"error:{type(exc).__name__}:{exc}")
summary.append(res)
if res["status"] in ("ok", "low-confidence"):
d = res["depth"]
flag = "" if res["confident"] else " <-- LOW CONFIDENCE (skipped, likely wrong room)"
print(f"{sid}: {res['scene_uuid'][:8]}/{res['room']:26} "
f"obj={res['n_objects']:2d} ann={res['n_annotated']} "
f"cover={d['coverage']*100:4.1f}% med={d['med_mm']:5.1f}mm "
f"<5mm={d['within5mm']*100:4.1f}%{flag}")
else:
print(f"{sid}: {res['status']}")
json.dump(summary, open(os.path.join(args.out, "gt_3dfront_summary.json"), "w"), indent=2)
if __name__ == "__main__":
main()