File size: 11,220 Bytes
460dc79 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 | """ARC-AGI-2 Video Answer Evaluator.
Extracts the test output grid from the last frame of a generated video,
then compares it against the ground-truth answer.
Color recovery pipeline:
1. Match pixel RGB against the canonical ARC_COLORS palette β permuted color index
2. Apply inverse permutation β original color index
3. Compare with ground truth
Usage:
python video_evaluate.py --video_dir videos --data_dir data --output results.json
"""
import json
import random
import argparse
from pathlib import Path
from collections import defaultdict
import cv2
import numpy as np
from tqdm import tqdm
# ββ ARC Color Palette (RGB) βββββββββββββββββββββββββββββββββββββββββββββββββββ
ARC_COLORS = np.array([
[0x00, 0x00, 0x00], # 0: black
[0x00, 0x74, 0xD9], # 1: blue
[0xFF, 0x41, 0x36], # 2: red
[0x2E, 0xCC, 0x40], # 3: green
[0xFF, 0xDC, 0x00], # 4: yellow
[0xAA, 0xAA, 0xAA], # 5: grey
[0xF0, 0x12, 0xBE], # 6: magenta
[0xFF, 0x85, 0x1B], # 7: orange
[0x7F, 0xDB, 0xFF], # 8: light blue
[0x87, 0x0C, 0x25], # 9: maroon
], dtype=np.uint8)
# ββ Color Permutation Utilities ββββββββββββββββββββββββββββββββββββββββββββββββ
def generate_color_permutation(seed: int) -> list[int]:
"""Reproduce the same permutation used during video generation."""
rng = random.Random(seed)
perm = list(range(10))
rng.shuffle(perm)
return perm
def invert_permutation(perm: list[int]) -> list[int]:
"""Compute inverse permutation: inv[perm[i]] = i."""
inv = [0] * len(perm)
for i, p in enumerate(perm):
inv[p] = i
return inv
# ββ Layout Computation (mirrors video_generate.py exactly) βββββββββββββββββββββ
def compute_test_output_bbox(task: dict, canvas_h: int, canvas_w: int) -> dict:
"""Compute pixel bounding box of the test output grid region.
Replicates _compute_layout + render_frame positioning from video_generate.py.
"""
n_cols = len(task["train"]) + 1
n_rows = 2
padding = 12
outer_margin = 16
label_h = 20
usable_w = canvas_w - 2 * outer_margin - (n_cols - 1) * padding
usable_h = canvas_h - 2 * outer_margin - (n_rows - 1) * padding
cell_w = usable_w // n_cols
cell_h = usable_h // n_rows
total_block_w = cell_w * n_cols + (n_cols - 1) * padding
total_block_h = cell_h * n_rows + (n_rows - 1) * padding
margin_x = (canvas_w - total_block_w) // 2
margin_y = (canvas_h - total_block_h) // 2
# Test output: last column, second row
col = n_cols - 1
x0 = margin_x + col * (cell_w + padding)
y0 = margin_y + cell_h + padding
test_out = np.array(task["test"][0]["output"])
gr, gc = test_out.shape
return {
"grid_rows": gr,
"grid_cols": gc,
"grid_x0": x0,
"grid_y0": y0 + label_h,
"grid_w": cell_w,
"grid_h": cell_h - label_h,
}
# ββ Frame Extraction βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def extract_last_frame(video_path: str) -> np.ndarray:
"""Extract the last frame from a video as an RGB numpy array."""
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
raise FileNotFoundError(f"Cannot open video: {video_path}")
total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
cap.set(cv2.CAP_PROP_POS_FRAMES, max(0, total - 1))
ret, frame = cap.read()
cap.release()
if not ret:
raise RuntimeError(f"Failed to read last frame from {video_path}")
return cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
# ββ Grid Extraction ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def extract_grid_from_frame(
frame: np.ndarray,
grid_x0: int,
grid_y0: int,
grid_w: int,
grid_h: int,
grid_rows: int,
grid_cols: int,
) -> list[list[int]]:
"""Extract ARC grid by sampling cell centers and matching to ARC_COLORS.
Always matches against the canonical ARC_COLORS palette. The returned
indices are the permuted color values as rendered in the video.
Args:
frame: RGB image (H, W, 3).
grid_x0, grid_y0: Top-left of grid area (below label).
grid_w, grid_h: Grid area dimensions.
grid_rows, grid_cols: Expected grid shape.
Returns:
Grid of permuted color indices (apply inverse perm to get originals).
"""
cell_h = grid_h / grid_rows
cell_w = grid_w / grid_cols
grid = []
for r in range(grid_rows):
row = []
cy = int(grid_y0 + (r + 0.5) * cell_h)
for c in range(grid_cols):
cx = int(grid_x0 + (c + 0.5) * cell_w)
# 3x3 patch average for codec artifact robustness
patch = frame[max(0, cy - 1): cy + 2, max(0, cx - 1): cx + 2]
avg = patch.mean(axis=(0, 1)).astype(np.uint8)
dists = np.sum((ARC_COLORS.astype(int) - avg.astype(int)) ** 2, axis=1)
row.append(int(np.argmin(dists)))
grid.append(row)
return grid
# ββ Evaluation βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def evaluate_video(
video_path: str,
task: dict,
perm: list[int],
canvas_h: int = 720,
canvas_w: int = 1280,
) -> dict:
"""Evaluate a single video against ground truth.
Pipeline:
1. Extract last frame (full answer revealed)
2. Locate test output region via layout math
3. Sample cell centers β match to ARC_COLORS β get permuted color indices
4. Apply inverse permutation β recover original color indices
5. Compare with ground truth
Returns:
Dict with 'correct', 'predicted_grid', 'ground_truth', 'pixel_accuracy'.
"""
frame = extract_last_frame(video_path)
bbox = compute_test_output_bbox(task, canvas_h, canvas_w)
# Step 1: extract permuted color indices from rendered pixels
permuted_grid = extract_grid_from_frame(frame, **bbox)
# Step 2: invert permutation to recover original values
inv = invert_permutation(perm)
predicted = [[inv[cell] for cell in row] for row in permuted_grid]
# Step 3: compare with ground truth
gt = task["test"][0]["output"]
correct = (predicted == gt)
gt_flat = [c for row in gt for c in row]
pred_flat = [c for row in predicted for c in row]
n_match = sum(a == b for a, b in zip(gt_flat, pred_flat))
pixel_acc = n_match / max(len(gt_flat), 1)
return {
"correct": correct,
"predicted_grid": predicted,
"ground_truth": gt,
"pixel_accuracy": pixel_acc,
}
# ββ Batch Evaluation βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def evaluate_all(
video_dir: str = "videos",
data_dir: str = "data",
output_file: str = "results.json",
) -> None:
"""Evaluate all videos against ground-truth tasks.
Recovers the color permutation from the seed in the filename
({task_id}_{seed}.mp4) using the same RNG as video_generate.py.
"""
video_path = Path(video_dir)
data_path = Path(data_dir)
# Build task file lookup
task_files: dict[str, Path] = {}
for subdir in ["training", "evaluation"]:
d = data_path / subdir
if d.exists():
for fp in d.glob("*.json"):
task_files[fp.stem] = fp
videos = sorted(video_path.glob("*.mp4"))
if not videos:
print(f"No videos found in {video_dir}")
return
# Auto-detect resolution from first video
cap = cv2.VideoCapture(str(videos[0]))
canvas_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
canvas_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
cap.release()
print(f"Detected resolution: {canvas_h}x{canvas_w}")
results = {}
total_correct = 0
total_count = 0
for vp in tqdm(videos, desc="Evaluating"):
stem = vp.stem
parts = stem.rsplit("_", 1)
if len(parts) != 2:
continue
task_id, seed_str = parts
if task_id not in task_files:
tqdm.write(f"Skip {stem}: task not found")
continue
with open(task_files[task_id]) as f:
task = json.load(f)
if not task.get("test") or "output" not in task["test"][0]:
continue
# Recover the exact permutation from seed
seed = int(seed_str)
perm = generate_color_permutation(seed)
try:
result = evaluate_video(str(vp), task, perm, canvas_h, canvas_w)
results[stem] = {
"correct": result["correct"],
"pixel_accuracy": result["pixel_accuracy"],
"task_id": task_id,
"seed": seed_str,
}
total_count += 1
if result["correct"]:
total_correct += 1
except Exception as e:
tqdm.write(f"Error {stem}: {e}")
results[stem] = {"error": str(e), "task_id": task_id}
acc = total_correct / max(total_count, 1)
# Per-task pixel accuracy aggregation
task_pixels: dict[str, list[float]] = defaultdict(list)
for v in results.values():
if "pixel_accuracy" in v:
task_pixels[v["task_id"]].append(v["pixel_accuracy"])
per_task_pixel_acc = {
tid: round(sum(accs) / len(accs), 4)
for tid, accs in sorted(task_pixels.items())
}
summary = {
"total_videos": total_count,
"correct": total_correct,
"accuracy": round(acc, 4),
"mean_pixel_accuracy": round(
sum(per_task_pixel_acc.values()) / max(len(per_task_pixel_acc), 1), 4
),
"per_task_pixel_accuracy": per_task_pixel_acc,
"results": results,
}
with open(output_file, "w") as f:
json.dump(summary, f, indent=2)
print(f"\nResults: {total_correct}/{total_count} correct ({acc:.2%})")
print(f"Mean pixel accuracy (per-task avg): {summary['mean_pixel_accuracy']:.2%}")
print(f"Saved to {output_file}")
# ββ CLI ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
if __name__ == "__main__":
p = argparse.ArgumentParser(description="ARC Video Evaluator")
p.add_argument("--video_dir", type=str, default="videos")
p.add_argument("--data_dir", type=str, default="data")
p.add_argument("--output", type=str, default="results.json")
args = p.parse_args()
evaluate_all(args.video_dir, args.data_dir, args.output) |