SpaVoBench / scripts /append_spavobench_release.py
Xinyun-Liu's picture
Add timestamped keep samples from 2026-08-10 review
3a9b9c4 verified
Raw
History Blame Contribute Delete
16.3 kB
#!/usr/bin/env python3
"""Append reproducible, timestamped additions from a newer review export."""
from __future__ import annotations
import argparse
import collections
import concurrent.futures
import io
import json
import os
import re
import shutil
import subprocess
from pathlib import Path
from typing import Any
from PIL import Image
from prepare_spavobench_release import (
HAN,
SOURCE_REPOSITORY,
build_manifest,
build_source_manifest,
imageio_ffmpeg,
normalize_english,
normalize_scene,
parse_timestamp,
write_json,
write_jsonl,
)
REVIEW_VERSION = "worldmodelbench-review-2026-08-10.json"
TIMESTAMP_AT_END = re.compile(r"\s*[((]\s*(\d{1,4}(?:\.\d+)?)\s*[))]\s*$")
NUMBERED_ITEM = re.compile(r"(?m)^\s*(\d+)\s*[.)]\s*")
# These translations correspond to newly accepted records in the 2026-08-10
# export. The keys are export indices so duplicate sample IDs stay distinct.
TRANSLATIONS_BY_SOURCE_INDEX: dict[int, list[str]] = {
6: ["Open the drawer, then close it."],
7: ["Spread the blue bedsheet on the bed."],
9: [
"Take out the blue bottle that the person's hand is touching.",
"Rotate the small rectangular wooden block and apply clear glue to its side.",
"Fit the small rectangular wooden block into the lower-left corner of the rectangle.",
],
10: ["Exit through the door and walk downstairs."],
13: ["Descend slowly along the rock wall."],
14: ["Rotate the pedal clockwise to turn the wheel.", "Rotate the pedal clockwise to turn the wheel."],
16: ["Exit the door.", "Ride the electric bike out through the door.", "Turn left at the intersection."],
17: [
"Run forward to receive a pass from the teammate in black, then shoot the ball into the goal while a player in white defends.",
"Run quickly forward to catch up with the soccer ball.",
"Avoid the defender and kick the ball into the goal.",
"Receive a forward pass from the teammate in black, break through past the defender and goalkeeper, and shoot.",
],
20: ["Remove the rear wheel from the bicycle."],
21: ["Run forward, catch the ball, jump, and complete a dunk while keeping the camera pointed at the hoop.", "Perform a jump shot."],
22: ["Rotate the camera to the left.", "Rotate the camera to the right."],
25: ["Dribble the ball alternately with the left and right hands. Identify the spatial motion during this basketball activity."],
31: ["Drive along the road until the turn is fully completed."],
34: ["Drive forward around the flower bed."],
45: ["Place the transparent inflatable bag into the cardboard box."],
47: ["Pick up the cup by its handle."],
51: ["Return the top item in the box to the shelf."],
52: ["Sweep the paper ball on the floor from near to far."],
53: ["Sweep the paper ball on the floor from far to near."],
56: ["Drive along the highway."],
60: ["Turn left at the intersection ahead."],
62: ["Drive along the highway."],
63: ["Drive along the road."],
72: ["Make a 180-degree turn, then drive forward."],
75: ["Cut the avocado with a knife and show it to the camera."],
80: ["The traffic light is red; continue driving for the next five seconds."],
137: ["Ride the skateboard from right to left in the U-shaped ramp."],
138: ["Circle around the building once."],
139: ["Roll up the sweet pastry in baking paper, covering the corners securely while it chills in the refrigerator."],
162: [
"Lay the carpet in the room.",
"Move the flower bed to the woman's left and remove the original flower bed.",
"Finally, put the magazines and remote control into the basket.",
],
163: ["Put the plate on the second rack of the oven and close the oven door."],
165: ["Hold the camera level and circle the dining table once to show the dining room."],
177: ["Skateboard up the steps while keeping the camera following the skateboard."],
179: ["Skateboard onto the handrail."],
208: ["Walk forward among the oncoming crowd."],
213: ["Move toward the camera to show the photo clearly."],
218: ["Walk forward."],
219: ["Rotate the camera to the right to show the other side of the room."],
221: ["Rotate the camera to the left."],
222: ["Rotate the camera to the left."],
224: ["Maintain distance from the vehicle ahead and drive along the road until the turn is fully completed."],
225: ["The man walks upstairs while the camera moves backward to maintain a constant distance from him."],
226: ["Move the camera to the right."],
227: ["Drive along the highway."],
229: ["Walk to the far side of the table while rotating the camera to the right, keeping the table centered in view."],
231: ["Walk to the left side of the table while rotating the camera to the right, keeping the table centered in view."],
}
# These accepted rows have a timestamp but their captions terminate in an
# incomplete word. They remain in the audit file until corrected at source.
INCOMPLETE_CAPTION_SOURCE_INDICES = {223, 228}
def source_key(row: dict[str, Any]) -> tuple[str, str]:
return str(row.get("sample_id") or ""), str(row.get("video_path") or "")
def split_caption_with_preamble(caption: str) -> list[dict[str, Any]]:
matches = list(NUMBERED_ITEM.finditer(caption))
if not matches:
chunks = [(None, caption)]
else:
chunks: list[tuple[int | None, str]] = []
preamble = caption[: matches[0].start()].strip()
if preamble:
chunks.append((None, preamble))
for position, match in enumerate(matches):
end = matches[position + 1].start() if position + 1 < len(matches) else len(caption)
chunks.append((int(match.group(1)), caption[match.end() : end]))
parts: list[dict[str, Any]] = []
for item_number, text in chunks:
text = text.strip()
timestamp = None
found = TIMESTAMP_AT_END.search(text)
if found:
timestamp = found.group(1)
text = text[: found.start()].strip()
parts.append({"source_item_number": item_number, "raw_caption": text, "timestamp_raw": timestamp})
return parts
def translated_caption(source_index: int, segment_index: int, raw_caption: str) -> str:
overrides = TRANSLATIONS_BY_SOURCE_INDEX.get(source_index)
if overrides is not None:
if segment_index > len(overrides):
raise ValueError(f"missing translation for source row {source_index}, segment {segment_index}")
caption = overrides[segment_index - 1]
else:
caption = raw_caption
caption = normalize_english(caption)
if HAN.search(caption):
raise ValueError(f"untranslated Chinese remains in source row {source_index}, segment {segment_index}")
return caption
def load_existing_records(release: Path) -> list[dict[str, Any]]:
payload = json.loads((release / "data" / "annotations.json").read_text(encoding="utf-8"))
return list(payload["annotations"])
def select_added_source_rows(old_source: dict[str, Any], new_source: dict[str, Any]) -> list[tuple[int, dict[str, Any]]]:
old_keep_keys = {source_key(row) for row in old_source["annotations"] if row.get("decision") == "keep"}
return [
(index, row)
for index, row in enumerate(new_source["annotations"])
if row.get("decision") == "keep" and source_key(row) not in old_keep_keys
]
def build_candidates(
source_rows: list[tuple[int, dict[str, Any]]],
existing: list[dict[str, Any]],
) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
known_source_indices = {
row.get("source_annotation_index")
for row in existing
if row.get("source_review_version") == REVIEW_VERSION
}
candidates: list[dict[str, Any]] = []
audit: list[dict[str, Any]] = []
for source_index, row in source_rows:
if source_index in known_source_indices:
continue
parts = split_caption_with_preamble(str(row.get("text_caption") or ""))
for segment_index, part in enumerate(parts, start=1):
raw_caption = part["raw_caption"]
if not raw_caption:
audit.append({"kind": "empty_caption", "source_annotation_index": source_index, "sample_id": row.get("sample_id")})
continue
if source_index in INCOMPLETE_CAPTION_SOURCE_INDICES:
audit.append({"kind": "incomplete_caption", "source_annotation_index": source_index, "sample_id": row.get("sample_id")})
continue
timestamp_raw = part["timestamp_raw"]
if not timestamp_raw and (len(parts) == 1 or part["source_item_number"] is None):
timestamp_raw = str(row.get("timestamp_seconds") or "").strip()
timestamp_seconds = parse_timestamp(timestamp_raw)
if timestamp_seconds is None:
audit.append(
{
"kind": "missing_timestamp",
"source_annotation_index": source_index,
"source_caption_segment_index": segment_index,
"sample_id": row.get("sample_id"),
}
)
continue
caption = translated_caption(source_index, segment_index, raw_caption)
annotation_id = f"spavobench-20260810-{source_index:04d}-{segment_index:02d}"
candidates.append(
{
"annotation_id": annotation_id,
"sample_id": row["sample_id"],
"source_annotation_index": source_index,
"source_review_version": REVIEW_VERSION,
"source_caption_segment_index": segment_index,
"source_caption_item_number": part["source_item_number"],
"text_caption": caption,
"caption_status": "complete",
"timestamp_raw": timestamp_raw,
"timestamp_seconds": timestamp_seconds,
"timestamp_provenance": "caption" if part["timestamp_raw"] else "record",
"data_source": row.get("data_source", ""),
"annotator": row.get("annotator", ""),
"decision": "keep",
"track": row.get("track", ""),
"scene": normalize_scene(row.get("scene", "")),
"spatial_ability": row.get("spatial_ability", ""),
"perspective": row.get("perspective", ""),
"video_path": row.get("video_path", ""),
"source_video_url": row.get("video_url", ""),
"source_video_repository": SOURCE_REPOSITORY,
"source_video_included": False,
"submitted_at": row.get("submitted_at"),
"local_saved_at": row.get("local_saved_at"),
"updated_at": row.get("updated_at"),
"frame_timestamp_seconds": timestamp_seconds,
"image_path": f"images/{annotation_id}.jpg",
}
)
return candidates, audit
def extract_candidates(candidates: list[dict[str, Any]], output: Path, workers: int, http_proxy: str) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
ffmpeg = imageio_ffmpeg.get_ffmpeg_exe()
image_root = output / "images"
audit: list[dict[str, Any]] = []
def run_frame_job(row: dict[str, Any]) -> tuple[str, str | None]:
destination = image_root / f"{row['annotation_id']}.jpg"
destination.parent.mkdir(parents=True, exist_ok=True)
if destination.exists() and destination.stat().st_size > 0:
return row["annotation_id"], None
command = [ffmpeg, "-hide_banner", "-loglevel", "error"]
if http_proxy:
command.extend(["-http_proxy", http_proxy])
command.extend(
[
"-ss",
f"{row['frame_timestamp_seconds']:.3f}",
"-i",
row["source_video_url"],
"-frames:v",
"1",
"-vf",
"format=yuv420p",
"-q:v",
"2",
str(destination),
]
)
try:
subprocess.run(command, check=True, timeout=240, capture_output=True)
if not destination.exists() or destination.stat().st_size == 0:
return row["annotation_id"], "ffmpeg completed without an image"
return row["annotation_id"], None
except subprocess.TimeoutExpired:
return row["annotation_id"], "remote frame extraction timed out after 240 seconds"
except subprocess.CalledProcessError as error:
fallback = command[:-2] + ["-f", "image2pipe", "-vcodec", "png", "-"]
try:
decoded = subprocess.run(fallback, check=True, timeout=240, capture_output=True).stdout
with Image.open(io.BytesIO(decoded)) as image:
image.convert("RGB").save(destination, format="JPEG", quality=95)
return row["annotation_id"], None
except Exception:
return row["annotation_id"], f"remote ffmpeg failed with exit status {error.returncode}"
failures: dict[str, str] = {}
with concurrent.futures.ThreadPoolExecutor(max_workers=workers) as pool:
for annotation_id, error in pool.map(run_frame_job, candidates):
if error:
failures[annotation_id] = error
for row in candidates:
if row["annotation_id"] in failures:
audit.append({"kind": "frame", "annotation_id": row["annotation_id"], "error": failures[row["annotation_id"]]})
return [row for row in candidates if row["annotation_id"] not in failures], audit
def append_release(old_source_path: Path, new_source_path: Path, output: Path, workers: int, http_proxy: str) -> dict[str, Any]:
old_source = json.loads(old_source_path.read_text(encoding="utf-8"))
new_source = json.loads(new_source_path.read_text(encoding="utf-8"))
existing = load_existing_records(output)
candidates, audit = build_candidates(select_added_source_rows(old_source, new_source), existing)
added, extraction_audit = extract_candidates(candidates, output, workers, http_proxy)
audit.extend(extraction_audit)
merged = existing + added
manifest = build_manifest(merged, audit)
manifest["incremental_update"] = {
"review_file": REVIEW_VERSION,
"source_keep_rows_added": len(select_added_source_rows(old_source, new_source)),
"image_caption_records_added": len(added),
"skipped_or_failed_records": len(audit),
}
write_json(output / "data" / "annotations.json", {"schema_version": "spavobench-v1", "annotations": merged})
write_jsonl(output / "data" / "annotations.jsonl", merged)
write_json(output / "data" / "manifest.json", manifest)
write_json(output / "data" / "update-2026-08-10.json", {"review_file": REVIEW_VERSION, "added_records": added, "audit": audit})
write_jsonl(output / "data" / "frame_failures.jsonl", audit)
write_jsonl(output / "source-videos" / "manifest.jsonl", build_source_manifest(merged))
(output / "scripts").mkdir(exist_ok=True)
shutil.copy2(Path(__file__), output / "scripts" / Path(__file__).name)
return manifest
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--old-source", type=Path, default=Path("/Users/liusmac/Downloads/worldmodelbench-review-2026-08-08 (6).json"))
parser.add_argument("--new-source", type=Path, default=Path("/Users/liusmac/Downloads/worldmodelbench-review-2026-08-10.json"))
parser.add_argument("--output", type=Path, default=Path("artifacts/spavobench/v1/release"))
parser.add_argument("--download-workers", type=int, default=4)
parser.add_argument("--http-proxy", default=os.environ.get("SPAVOBENCH_HTTP_PROXY", "http://127.0.0.1:7897"))
args = parser.parse_args()
print(json.dumps(append_release(args.old_source, args.new_source, args.output, args.download_workers, args.http_proxy), ensure_ascii=False, indent=2))
if __name__ == "__main__":
main()