File size: 11,164 Bytes
3726258 3f3fbe2 3726258 3f3fbe2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 | """Deterministic layout solver: one box per class from per-class size priors (``class_size_priors.json``).
Used for k = 5 scenes, for never-co-occurring cliques and as the fallback of the language-model planner.
"""
from __future__ import annotations
import json
import math
from dataclasses import dataclass
from pathlib import Path
from typing import Dict, List, Optional, Sequence, Tuple
import numpy as np
DEFAULT_PRIORS_PATH = Path(__file__).resolve().parent / "class_size_priors.json"
DEFAULT_PRIOR: Dict[str, float] = {
"rel_area_p25": 0.03, "rel_area_p50": 0.08, "rel_area_p75": 0.18,
"aspect_median": 1.0,
}
VEHICLES = {"bicycle", "car", "motorcycle", "bus", "train", "truck", "boat"}
FURNITURE = {"bench", "chair", "couch", "bed", "dining table", "toilet"}
ANIMALS = {"bird", "cat", "dog", "horse", "sheep", "cow", "elephant", "bear",
"zebra", "giraffe"}
PEOPLE = {"person"}
GROUND_CONTACT_CLASSES = VEHICLES | FURNITURE | ANIMALS | PEOPLE
@dataclass
class SolverConfig:
img_size: int = 1024
margin: int = 24
min_short_side: int = 96
min_rel_area: float = 0.02
max_rel_area: float = 0.30
max_iou: float = 0.02
min_center_dist_frac: float = 0.25
max_attempts: int = 2000
shrink_factor: float = 0.85
max_shrink_rounds: int = 12
ground_contact_lower_frac: float = 0.60
area_sigma_scale: float = 1.0
area_scale: float = 1.0
aspect_sigma: float = 0.15
min_aspect: float = 0.33
max_aspect: float = 3.0
def load_size_priors(path: Path = DEFAULT_PRIORS_PATH) -> Dict[str, Dict[str, float]]:
with open(path) as f:
d = json.load(f)
return d.get("classes", d)
def box_iou(a: Sequence[float], b: Sequence[float]) -> float:
ix1, iy1 = max(a[0], b[0]), max(a[1], b[1])
ix2, iy2 = min(a[2], b[2]), min(a[3], b[3])
iw, ih = max(0.0, ix2 - ix1), max(0.0, iy2 - iy1)
inter = iw * ih
if inter <= 0:
return 0.0
ua = (a[2] - a[0]) * (a[3] - a[1]) + (b[2] - b[0]) * (b[3] - b[1]) - inter
return inter / ua if ua > 0 else 0.0
def _center(b: Sequence[float]) -> Tuple[float, float]:
return (b[0] + b[2]) / 2.0, (b[1] + b[3]) / 2.0
def _diag(b: Sequence[float]) -> float:
return math.hypot(b[2] - b[0], b[3] - b[1])
def pair_ok(a: Sequence[float], b: Sequence[float], cfg: SolverConfig) -> bool:
if box_iou(a, b) > cfg.max_iou:
return False
ca, cb = _center(a), _center(b)
dist = math.hypot(ca[0] - cb[0], ca[1] - cb[1])
return dist >= cfg.min_center_dist_frac * min(_diag(a), _diag(b))
def _apply_size_floors(w: float, h: float, cfg: SolverConfig) -> Tuple[float, float]:
"""Short side >= min_short_side, area <= max_rel_area, inside the canvas."""
canvas = float(cfg.img_size)
max_side = canvas - 2 * cfg.margin
if min(w, h) < cfg.min_short_side:
if w <= h:
w = float(cfg.min_short_side)
else:
h = float(cfg.min_short_side)
cap = cfg.max_rel_area * canvas * canvas
if w * h > cap:
s = math.sqrt(cap / (w * h))
w, h = max(w * s, cfg.min_short_side), max(h * s, cfg.min_short_side)
w, h = min(w, max_side), min(h, max_side)
return w, h
def sample_target_sizes(labels: Sequence[str], priors: Dict[str, Dict[str, float]],
rng: np.random.Generator, cfg: SolverConfig
) -> List[Tuple[float, float, float]]:
"""Per label: (w, h, target_rel_area) before any shrink round."""
canvas = float(cfg.img_size)
out = []
for lbl in labels:
p = priors.get(lbl, DEFAULT_PRIOR)
p25 = max(float(p.get("rel_area_p25", DEFAULT_PRIOR["rel_area_p25"])), 1e-4)
p50 = max(float(p.get("rel_area_p50", DEFAULT_PRIOR["rel_area_p50"])), 1e-4)
p75 = max(float(p.get("rel_area_p75", DEFAULT_PRIOR["rel_area_p75"])), p25)
sigma = cfg.area_sigma_scale * (math.log(p75) - math.log(p25)) / 1.349
rel = math.exp(math.log(p50) + rng.normal(0.0, max(sigma, 1e-3)))
rel = min(max(rel, p25), p75) * cfg.area_scale
rel = min(max(rel, cfg.min_rel_area), cfg.max_rel_area)
ar = float(p.get("aspect_median", 1.0)) * math.exp(rng.normal(0.0, cfg.aspect_sigma))
ar = min(max(ar, cfg.min_aspect), cfg.max_aspect)
area_px = rel * canvas * canvas
w = math.sqrt(area_px * ar)
h = math.sqrt(area_px / ar)
w, h = _apply_size_floors(w, h, cfg)
out.append((w, h, rel))
return out
def _try_place(order: List[int], dims: List[Tuple[float, float]],
labels: Sequence[str], rng: np.random.Generator,
cfg: SolverConfig, attempt_budget: int
) -> Tuple[Optional[List[List[int]]], int]:
"""One placement pass. Returns (boxes in label order or None, attempts used)."""
canvas = cfg.img_size
placed: Dict[int, List[int]] = {}
attempts = 0
per_obj = max(1, attempt_budget // max(1, len(order)))
for i in order:
w, h = dims[i]
wi, hi = int(math.floor(w)), int(math.floor(h))
x_lo, x_hi = cfg.margin, canvas - cfg.margin - wi
if labels[i] in GROUND_CONTACT_CLASSES:
y2_lo = max(cfg.margin + hi,
int(round((1.0 - cfg.ground_contact_lower_frac) * canvas)))
y2_hi = canvas - cfg.margin
y_lo, y_hi = y2_lo - hi, y2_hi - hi
else:
y_lo, y_hi = cfg.margin, canvas - cfg.margin - hi
if x_hi < x_lo or y_hi < y_lo:
return None, attempts
ok = False
for _ in range(per_obj):
attempts += 1
x1 = int(rng.integers(x_lo, x_hi + 1))
y1 = int(rng.integers(y_lo, y_hi + 1))
cand = [x1, y1, x1 + wi, y1 + hi]
if all(pair_ok(cand, b, cfg) for b in placed.values()):
placed[i] = cand
ok = True
break
if not ok:
return None, attempts
return [placed[i] for i in range(len(labels))], attempts
def solve_layout(labels: Sequence[str], priors: Dict[str, Dict[str, float]],
img_size: int = 1024, seed: int = 0,
cfg: Optional[SolverConfig] = None, area_scale: float = 1.0) -> Dict:
"""Return ``{layout_boxes, layout_source, layout_attempts, layout_shrink_rounds,
target_rel_areas}``. ``layout_source`` is ``"solver"`` on success and
``"grid_fallback"`` only if every shrink round failed (not expected for
<= 8 objects at the configured floors). ``layout_attempts`` counts solver
placements, not LLM calls; Stage 1 renames it to ``solver_attempts``."""
cfg = cfg or SolverConfig()
if cfg.img_size != img_size or cfg.area_scale != area_scale:
cfg = SolverConfig(**{**cfg.__dict__, "img_size": img_size,
"area_scale": float(area_scale)})
rng = np.random.default_rng(int(seed))
sizes = sample_target_sizes(labels, priors, rng, cfg)
base_dims = [(w, h) for w, h, _ in sizes]
areas = [w * h for w, h in base_dims]
order = sorted(range(len(labels)), key=lambda i: -areas[i])
total_attempts = 0
for rnd in range(cfg.max_shrink_rounds + 1):
s = cfg.shrink_factor ** rnd
dims = [_apply_size_floors(w * s, h * s, cfg) for w, h in base_dims]
boxes, used = _try_place(order, dims, labels, rng, cfg, cfg.max_attempts)
total_attempts += used
if boxes is not None:
return {
"layout_boxes": boxes,
"layout_source": "solver",
"layout_attempts": total_attempts,
"layout_shrink_rounds": rnd,
"target_rel_areas": [round(float(r), 5) for _, _, r in sizes],
}
from strangebench.scene_specs.select_scenes_and_layouts import _grid_fallback
return {
"layout_boxes": _grid_fallback(list(labels), img_size),
"layout_source": "grid_fallback",
"layout_attempts": total_attempts,
"layout_shrink_rounds": cfg.max_shrink_rounds,
"target_rel_areas": [round(float(r), 5) for _, _, r in sizes],
}
def check_layout(boxes: Sequence[Sequence[int]], labels: Sequence[str],
cfg: Optional[SolverConfig] = None) -> List[str]:
"""Return a list of violated rules (empty = valid). Used by the tests and
by the Stage 1 driver as a cheap assertion on solver output."""
cfg = cfg or SolverConfig()
canvas = cfg.img_size
problems: List[str] = []
for i, b in enumerate(boxes):
x1, y1, x2, y2 = b
if not (cfg.margin <= x1 < x2 <= canvas - cfg.margin
and cfg.margin <= y1 < y2 <= canvas - cfg.margin):
problems.append(f"{labels[i]}: outside margins {b}")
if min(x2 - x1, y2 - y1) < cfg.min_short_side:
problems.append(f"{labels[i]}: short side < {cfg.min_short_side}")
rel = (x2 - x1) * (y2 - y1) / float(canvas * canvas)
if rel > cfg.max_rel_area + 1e-6:
problems.append(f"{labels[i]}: rel area {rel:.3f} > {cfg.max_rel_area}")
if labels[i] in GROUND_CONTACT_CLASSES and \
y2 < (1.0 - cfg.ground_contact_lower_frac) * canvas - 1:
problems.append(f"{labels[i]}: ground-contact class floats (y2={y2})")
for i in range(len(boxes)):
for j in range(i + 1, len(boxes)):
if not pair_ok(boxes[i], boxes[j], cfg):
problems.append(f"{labels[i]} vs {labels[j]}: overlap/center rule")
return problems
def shrink_anchor_box(box: Sequence[int], img_size: int = 1024,
max_rel_area: float = 0.02, min_short_side: int = 72,
margin: int = 24) -> Tuple[List[int], float]:
"""Shrink ``box`` around its center so its area is <= ``max_rel_area`` of
the canvas while keeping the short side >= ``min_short_side``. The
aspect ratio is preserved unless the short-side floor forces the long
side down to keep the area cap. Returns (new_box, realized_rel_area)."""
x1, y1, x2, y2 = (float(v) for v in box)
w, h = x2 - x1, y2 - y1
cx, cy = (x1 + x2) / 2.0, (y1 + y2) / 2.0
cap = max_rel_area * img_size * img_size
if w * h > cap:
s = math.sqrt(cap / (w * h))
w, h = w * s, h * s
if min(w, h) < min_short_side:
if w <= h:
w = float(min_short_side)
h = min(h, cap / w)
else:
h = float(min_short_side)
w = min(w, cap / h)
w, h = max(w, min(min_short_side, img_size - 2 * margin)), \
max(h, min(min_short_side, img_size - 2 * margin))
wi, hi = int(math.floor(w)), int(math.floor(h))
nx1, ny1 = cx - wi / 2.0, cy - hi / 2.0
nx1 = min(max(nx1, margin), img_size - margin - wi)
ny1 = min(max(ny1, margin), img_size - margin - hi)
x1i, y1i = int(round(nx1)), int(round(ny1))
nb = [x1i, y1i, x1i + wi, y1i + hi]
rel = (nb[2] - nb[0]) * (nb[3] - nb[1]) / float(img_size * img_size)
return nb, rel
def rel_area_of_box(box: Sequence[float], img_size: int = 1024) -> float:
return max(0.0, (box[2] - box[0])) * max(0.0, (box[3] - box[1])) / float(img_size * img_size)
|