activevision's picture
Initial release v0.4.0 — ActiveVision benchmark (85 instances, 17 tasks)
f69e256 verified
raw
history blame
21.9 kB
"""Generate constellation_match_count samples.
Each sample produces a SINGLE side-by-side image: Template on the left,
Field on the right, separated by a thin divider.
The Template panel shows the reference constellation, and the Field shows
30-60 white "star" dots (with planted copies of the template among
distractor stars).
The task: count how many translated copies of the template pattern occur
in the Field (same relative offsets within a tolerance, no rotation/reflection).
The generator performs rejection sampling in TWO directions:
* it places the desired number of planted template instances, and
* it scatters additional non-template "distractor" stars while rejecting
any configuration that would accidentally form another template match.
After all dots are placed the final match-count is verified; a mismatch
against the intended count causes the sample to be regenerated.
"""
from __future__ import annotations
import argparse
import io
import json
import math
import random
from pathlib import Path
from typing import List, Tuple
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
from tqdm import tqdm
QUESTION = (
"This image has two panels separated by a thin vertical divider. "
"The left panel shows the Template: a small constellation of white dots. "
"The right panel shows the Field: a larger scene with many white stars. "
"The two panels are drawn at the SAME pixel scale and the dots are the "
"SAME pixel size. Count how many copies of the Template pattern appear "
"in the Field. A copy may be rotated by a small angle (up to about 20 "
"degrees in either direction) and individual dot positions may be jittered "
"slightly, but the overall pattern of relative dot positions must be "
"preserved. The patterns may be rotated by a small angle. No mirroring "
"or scaling. Report the count as a non-negative integer. "
"Provide your final answer enclosed in <answer>...</answer> tags."
)
# Per-copy rotation angle drawn from [-MAX_ANGLE_DEG, +MAX_ANGLE_DEG].
# Chosen large enough to break translation-only Hausdorff matching (which
# succeeded at 100% on the unrotated version) but small enough that human
# viewers readily identify each copy as the same pattern.
MAX_ANGLE_DEG = 20.0
# Candidate angles sampled in the matching verifier.
ANGLE_SEARCH_STEP = 2.0
# Shared dot size (matplotlib scatter ``s`` is marker area in points^2).
# Same value used for both the Template panel and the Field so the model has
# a direct pixel-to-pixel correspondence.
DOT_SIZE = 130
# ---------------------------------------------------------------------------
# Template generation
# ---------------------------------------------------------------------------
def random_template(rng: random.Random) -> np.ndarray:
"""Build a template of exactly 5 dots with distinct relative offsets.
Returned as an (n, 2) ndarray of offsets relative to the first dot,
so the first row is always (0, 0).
"""
n = 5
# Draw points on a coarse integer grid so that offsets are well separated,
# then jitter slightly to keep the pattern visually non-lattice.
scale = rng.uniform(36.0, 54.0) # pixel scale of the template
while True:
grid_pts = set()
grid_pts.add((0, 0))
while len(grid_pts) < n:
gx = rng.randint(-3, 3)
gy = rng.randint(-3, 3)
grid_pts.add((gx, gy))
pts_grid = list(grid_pts)
pts = []
for gx, gy in pts_grid:
jitter_x = rng.uniform(-4.0, 4.0)
jitter_y = rng.uniform(-4.0, 4.0)
pts.append((gx * scale + jitter_x, gy * scale + jitter_y))
arr = np.asarray(pts, dtype=np.float64)
# Normalise so first point is origin
arr = arr - arr[0]
# Reject degenerate patterns: ensure min pairwise distance is comfortable,
# and that the pattern occupies a reasonable bounding box.
diffs = arr[:, None, :] - arr[None, :, :]
dists = np.linalg.norm(diffs, axis=-1)
np.fill_diagonal(dists, np.inf)
if dists.min() < 28.0:
continue
bbox = arr.max(axis=0) - arr.min(axis=0)
if bbox[0] < 40.0 or bbox[1] < 40.0:
continue
if bbox[0] > 260.0 or bbox[1] > 260.0:
continue
return arr
# ---------------------------------------------------------------------------
# Template matching
# ---------------------------------------------------------------------------
def _rot_matrix(theta_deg: float) -> np.ndarray:
t = np.deg2rad(theta_deg)
c, s = np.cos(t), np.sin(t)
return np.array([[c, -s], [s, c]])
def count_template_matches(stars: np.ndarray, template: np.ndarray,
tol: float,
max_angle_deg: float = MAX_ANGLE_DEG,
angle_step_deg: float = ANGLE_SEARCH_STEP) -> int:
"""Count distinct (translation, rotation) placements such that every
(R(theta) @ template[i] + T) has a star within `tol` pixels, where T is
a candidate translation and theta is a rotation angle in
[-max_angle_deg, +max_angle_deg]. We deduplicate by translation only:
two matches are the same if their anchor translations differ by less
than `tol`.
stars: (S, 2) array, template: (n, 2) with template[0] = (0,0).
"""
if len(stars) == 0:
return 0
n = template.shape[0]
tol_sq = tol * tol
angles = np.arange(-max_angle_deg, max_angle_deg + 1e-6, angle_step_deg)
found_translations: List[np.ndarray] = []
for anchor in stars:
anchor_matched = False
for theta in angles:
R = _rot_matrix(float(theta))
rotated = template @ R.T # (n, 2), first row still (0, 0)
ok = True
for k in range(1, n):
target = anchor + rotated[k]
d2 = np.sum((stars - target) ** 2, axis=1)
if d2.min() > tol_sq:
ok = False
break
if ok:
anchor_matched = True
break
if not anchor_matched:
continue
# Deduplicate by anchor translation
is_new = True
for t in found_translations:
if np.sum((t - anchor) ** 2) < tol_sq:
is_new = False
break
if is_new:
found_translations.append(anchor.copy())
return len(found_translations)
# ---------------------------------------------------------------------------
# Sample building
# ---------------------------------------------------------------------------
def build_sample(rng: random.Random, width: int, height: int,
target_matches: int, total_stars: int,
tol: float) -> Tuple[np.ndarray, np.ndarray, int]:
"""Return (stars, template, realised_matches). Raises RuntimeError if
it could not construct the intended configuration after many tries."""
for attempt in range(30):
template = random_template(rng)
n_tmpl = template.shape[0]
# Margins for anchor placement so full template stays inside
tmin = template.min(axis=0)
tmax = template.max(axis=0)
pad = 60.0
anchor_xmin = pad - tmin[0]
anchor_xmax = width - pad - tmax[0]
anchor_ymin = pad - tmin[1]
anchor_ymax = height - pad - tmax[1]
if anchor_xmax - anchor_xmin < 100 or anchor_ymax - anchor_ymin < 100:
continue
placed_stars: List[np.ndarray] = []
# Separation so planted instances don't overlap each other excessively,
# and so the template anchors are themselves far enough apart to be
# counted as distinct translations.
min_anchor_sep = float(np.linalg.norm(tmax - tmin)) + 40.0
# 1. Plant template instances. Each planted copy carries a per-copy
# rotation in [-MAX_ANGLE_DEG, +MAX_ANGLE_DEG]. This breaks
# translation-only matching (Hausdorff + centroid-anchored offsets).
anchors: List[np.ndarray] = []
planted_angles: List[float] = []
for _ in range(target_matches):
ok = False
for _try in range(400):
ax = rng.uniform(anchor_xmin, anchor_xmax)
ay = rng.uniform(anchor_ymin, anchor_ymax)
theta = rng.uniform(-MAX_ANGLE_DEG, MAX_ANGLE_DEG)
candidate = np.array([ax, ay])
too_close = False
for a in anchors:
if np.linalg.norm(candidate - a) < min_anchor_sep:
too_close = True
break
if too_close:
continue
R = _rot_matrix(theta)
rotated_template = template @ R.T
# Check rotated bounding box still inside the panel padding.
rtmin = rotated_template.min(axis=0)
rtmax = rotated_template.max(axis=0)
if (candidate[0] + rtmin[0] < pad or
candidate[0] + rtmax[0] > width - pad or
candidate[1] + rtmin[1] < pad or
candidate[1] + rtmax[1] > height - pad):
continue
# Try adding this instance's dots while ensuring each new dot
# is at least `tol * 2.5` from all existing dots.
new_dots = [candidate + rotated_template[k] for k in range(n_tmpl)]
clash = False
for nd in new_dots:
for ex in placed_stars:
if np.linalg.norm(nd - ex) < tol * 2.5:
clash = True
break
if clash:
break
if clash:
continue
anchors.append(candidate)
planted_angles.append(float(theta))
placed_stars.extend(new_dots)
ok = True
break
if not ok:
break
if len(anchors) != target_matches:
continue
# 2. Scatter distractor stars, rejection-sample to avoid new matches
distractors_needed = total_stars - len(placed_stars)
if distractors_needed < 0:
continue
fail = False
for _d in range(distractors_needed):
added = False
for _try in range(500):
dx = rng.uniform(pad, width - pad)
dy = rng.uniform(pad, height - pad)
cand = np.array([dx, dy])
# Keep distractors from overlapping existing dots
too_close = False
for ex in placed_stars:
if np.linalg.norm(cand - ex) < 22.0:
too_close = True
break
if too_close:
continue
# Tentatively add and verify match count is unchanged
trial = np.asarray(placed_stars + [cand])
count = count_template_matches(trial, template, tol)
if count != target_matches:
continue
placed_stars.append(cand)
added = True
break
if not added:
fail = True
break
if fail:
continue
final_stars = np.asarray(placed_stars)
realised = count_template_matches(final_stars, template, tol)
if realised == target_matches:
return final_stars, template, realised
raise RuntimeError("Failed to build sample after many attempts")
# ---------------------------------------------------------------------------
# Rendering - two separate images per sample
# ---------------------------------------------------------------------------
def _add_dust(ax, rng: random.Random, width: int, height: int,
n_dust: int = 800) -> None:
dust_rng = np.random.default_rng(rng.randint(0, 2**31 - 1))
dx = dust_rng.uniform(0, width, size=n_dust)
dy = dust_rng.uniform(0, height, size=n_dust)
ds = dust_rng.uniform(0.3, 2.5, size=n_dust)
ax.scatter(dx, dy, s=ds, c="#1a2238", alpha=0.45, linewidths=0, zorder=1)
def render_field(width: int, height: int,
stars: np.ndarray, rng: random.Random) -> Image.Image:
"""Render the field image: dark sky with scattered white stars, no
template overlay. Returns a PIL Image."""
fig = plt.figure(figsize=(width / 100, height / 100), dpi=100,
facecolor="#0a1020")
ax = fig.add_axes([0, 0, 1, 1])
ax.set_xlim(0, width)
ax.set_ylim(height, 0)
ax.axis("off")
ax.set_facecolor("#0a1020")
_add_dust(ax, rng, width, height)
# All stars at a single, larger pixel size so the model can compare
# dot-to-dot spacings directly between Template and Field.
ax.scatter(stars[:, 0], stars[:, 1], s=DOT_SIZE, c="white", alpha=1.0,
linewidths=0, zorder=3)
buf = io.BytesIO()
fig.savefig(buf, format="png", dpi=100, bbox_inches="tight", pad_inches=0,
facecolor=fig.get_facecolor())
plt.close(fig)
buf.seek(0)
return Image.open(buf).convert("RGB")
def render_template(template: np.ndarray,
rng: random.Random) -> Image.Image:
"""Render the template panel at the SAME pixel scale as the field.
The canvas is a tight bounding box around the template dots plus a
small margin, so pixel-distances between template dots equal pixel-
distances between the corresponding dots in the field (1:1 scale).
A short header strip labels the panel as TEMPLATE.
Returns a PIL Image.
"""
from matplotlib.patches import Rectangle
tpl_min = template.min(axis=0)
tpl_max = template.max(axis=0)
span_x = float(tpl_max[0] - tpl_min[0])
span_y = float(tpl_max[1] - tpl_min[1])
margin = 50.0 # empty space around the dots
header_h = 34.0 # top strip for the TEMPLATE label
min_width = 240.0 # keep the header legible for narrow patterns
content_w = span_x + 2 * margin
content_h = span_y + 2 * margin
canvas_w = max(content_w, min_width)
canvas_h = header_h + content_h
# Offset that places tpl_min at (margin_x, header_h + margin) with the
# dots horizontally centred inside the canvas.
margin_x = (canvas_w - span_x) / 2.0
ox = margin_x - tpl_min[0]
oy = (header_h + margin) - tpl_min[1]
disp = template + np.array([ox, oy])
fig = plt.figure(figsize=(canvas_w / 100, canvas_h / 100), dpi=100,
facecolor="#070b14")
ax = fig.add_axes([0, 0, 1, 1])
ax.set_xlim(0, canvas_w)
ax.set_ylim(canvas_h, 0)
ax.axis("off")
ax.set_facecolor("#070b14")
_add_dust(ax, rng, int(canvas_w), int(canvas_h),
n_dust=max(40, int(canvas_w * canvas_h / 1800)))
# Header strip
ax.add_patch(Rectangle((0, 0), canvas_w, header_h,
facecolor="#11182a", edgecolor="none", zorder=4))
ax.plot([0, canvas_w], [header_h, header_h],
color="#2d3a5a", linewidth=1.0, zorder=5)
ax.text(canvas_w / 2, header_h / 2, "TEMPLATE",
color="#cfe0ff", fontsize=14, fontweight="bold",
ha="center", va="center", zorder=6,
family="DejaVu Sans")
# Subtle border around the whole canvas
ax.add_patch(Rectangle((1.5, 1.5), canvas_w - 3, canvas_h - 3,
facecolor="none", edgecolor="#7aa6ff",
linewidth=2.0, zorder=6))
ax.scatter(disp[:, 0], disp[:, 1], s=DOT_SIZE, c="white",
linewidths=0, zorder=7)
buf = io.BytesIO()
fig.savefig(buf, format="png", dpi=100, bbox_inches=None, pad_inches=0,
facecolor=fig.get_facecolor())
plt.close(fig)
buf.seek(0)
return Image.open(buf).convert("RGB")
BG_COLOR = "#0a1020"
DIVIDER_WIDTH = 3
DIVIDER_COLOR = (51, 51, 85) # #333355
def render_combined(out_path: Path, template_img: Image.Image,
field_img: Image.Image) -> None:
"""Concatenate template (left) and field (right) with a thin divider."""
bg = tuple(int(BG_COLOR.lstrip("#")[i:i+2], 16) for i in (0, 2, 4))
tw, th = template_img.size
fw, fh = field_img.size
combined_h = max(th, fh)
combined_w = tw + DIVIDER_WIDTH + fw
combined = Image.new("RGB", (combined_w, combined_h), bg)
# Vertically centre template
t_y = (combined_h - th) // 2
combined.paste(template_img, (0, t_y))
# Divider
for x in range(tw, tw + DIVIDER_WIDTH):
for y in range(combined_h):
combined.putpixel((x, y), DIVIDER_COLOR)
# Field
f_y = (combined_h - fh) // 2
combined.paste(field_img, (tw + DIVIDER_WIDTH, f_y))
combined.save(out_path)
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--output-root", type=Path, required=True)
parser.add_argument("--count", type=int, default=30)
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--width", type=int, default=900)
parser.add_argument("--height", type=int, default=900)
parser.add_argument("--tolerance", type=float, default=10.0,
help="Position tolerance (pixels) for a template match")
parser.add_argument("--difficulty", type=int, default=5,
help="Integer difficulty >=0; scales copies, angle, field stars.")
args = parser.parse_args()
d = max(0, int(args.difficulty))
# Canvas scaling: N_d = 50 + 5*d, N_0 = 50
N_d = 50 + 5 * d
N_0 = 50
s = math.sqrt(max(1.0, N_d / N_0))
args.width = int(round(args.width * s))
args.height = int(round(args.height * s))
if d > 0:
_min_copies = 5
_max_copies = 5 + d
# Override module-level MAX_ANGLE_DEG so both the builder and the
# verifier see the same scaled max angle.
global MAX_ANGLE_DEG
MAX_ANGLE_DEG = float(min(10, d))
_field_stars_target = 50 + 5 * d
else:
_min_copies = 5
_max_copies = 5
_field_stars_target = 50
out_root: Path = args.output_root
img_dir = out_root / "images"
img_dir.mkdir(parents=True, exist_ok=True)
ann_path = out_root / "annotations.jsonl"
master_rng = random.Random(args.seed)
# Force evenly-spaced answers across [_min_copies, _max_copies].
if args.count > 1:
plan = [
int(round(_min_copies + i * (_max_copies - _min_copies) / (args.count - 1)))
for i in range(args.count)
]
else:
plan = [_min_copies]
print(f"forced constellation match counts: {plan}")
records = []
with ann_path.open("w") as f:
for i in tqdm(range(args.count), desc="constellation_match_count"):
target = plan[i]
# Total star count: default 30-60, scaled a bit with target so high-match
# images stay legible. With difficulty override, use _field_stars_target.
if _field_stars_target is not None:
base_total = master_rng.randint(_field_stars_target,
_field_stars_target + 15)
total_stars = max(base_total, target * 5 + master_rng.randint(8, 18))
else:
base_total = master_rng.randint(30, 55)
total_stars = max(base_total, target * 5 + master_rng.randint(8, 18))
total_stars = min(total_stars, 60)
# Retry loop in case build_sample can't meet constraints
built = False
for retry in range(12):
sub_seed = master_rng.randint(0, 2**31 - 1)
sub_rng = random.Random(sub_seed)
try:
stars, template, realised = build_sample(
sub_rng, args.width, args.height,
target_matches=target,
total_stars=total_stars,
tol=args.tolerance,
)
except RuntimeError:
continue
built = True
break
if not built:
raise RuntimeError(f"sample {i} could not be generated")
img_name = f"constellation_match_count_{i:05d}.png"
tpl_img = render_template(template, random.Random(sub_seed + 2))
fld_img = render_field(args.width, args.height,
stars, random.Random(sub_seed + 1))
render_combined(img_dir / img_name, tpl_img, fld_img)
rec = {
"image": f"images/{img_name}",
"question": QUESTION,
"answer": realised,
"num_template_dots": int(template.shape[0]),
"total_stars": int(stars.shape[0]),
"num_matches": realised,
"metadata": {
"seed": sub_seed,
"tolerance": args.tolerance,
"template_offsets": template.tolist(),
},
}
f.write(json.dumps(rec) + "\n")
f.flush()
records.append(rec)
data_json = {
"task": "constellation_match_count",
"category": "visual_attribute_transfer",
"count": len(records),
"items": records,
}
(out_root / "data.json").write_text(json.dumps(data_json, indent=2))
print(f"Saved to {out_root}")
if __name__ == "__main__":
main()