Datasets:

vLAR
/

PhysInOne

"""Convert Unreal Movie Render Queue EXRs into PhysInOne dataset files.

RGB passes become JPEG files. World-depth and custom-stencil passes become
compressed NPZ files with keys ``depth`` and ``seg``. By default it also
builds ``points3d.ply`` from the first frame of every static camera. This tool
never samples views and never creates or modifies transforms_train/val/test.json.
"""

from __future__ import annotations

import argparse
from concurrent.futures import ThreadPoolExecutor, as_completed
from dataclasses import dataclass
import json
from pathlib import Path
import re
import shutil
import sys
import tempfile

import imageio.v2 as imageio
import Imath
import numpy as np
import OpenEXR

PACKAGE_ROOT = Path(__file__).resolve().parent.parent
if str(PACKAGE_ROOT) not in sys.path:
    sys.path.insert(0, str(PACKAGE_ROOT))

from postprocess.point_cloud import generate_points3d

try:
    from tqdm import tqdm
except ImportError:  # pragma: no cover
    def tqdm(iterable=None, **_kwargs):
        return iterable if iterable is not None else ()


PASS_PATTERNS = {
    "depth": re.compile(
        r"(?:FinalImage|PathTracer)MovieRenderQueue_WorldDepth_meter\.(\d+)\.exr$",
        re.IGNORECASE,
    ),
    "seg": re.compile(
        r"(?:FinalImage|PathTracer)CustomStencil_woText\.(\d+)\.exr$",
        re.IGNORECASE,
    ),
    "rgb": re.compile(r"(?:FinalImage|PathTracer)\.(\d+)\.exr$", re.IGNORECASE),
}


@dataclass(frozen=True)
class Job:
    source: Path
    destination: Path
    channel: str


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("render_dirs", nargs="+", type=Path)
    parser.add_argument(
        "--output-dir",
        type=Path,
        default=None,
        help="Optional destination for one input directory. Default: convert in place.",
    )
    parser.add_argument("--workers", type=int, default=8)
    parser.add_argument("--jpeg-quality", type=int, default=95)
    parser.add_argument(
        "--channels",
        default="rgb,depth,seg",
        help="Comma-separated subset of rgb,depth,seg.",
    )
    parser.add_argument(
        "--delete-source-exr",
        action="store_true",
        help="Delete each EXR only after its converted output passes validation.",
    )
    parser.add_argument(
        "--skip-ply",
        action="store_true",
        help="Do not generate points3d.ply after conversion.",
    )
    parser.add_argument(
        "--depth-threshold",
        type=float,
        default=20.0,
        help="Maximum valid depth in meters for points3d.ply (default: 20).",
    )
    parser.add_argument(
        "--max-points",
        type=int,
        default=100_000,
        help="Maximum number of vertices in points3d.ply (default: 100000).",
    )
    parser.add_argument("--ply-seed", type=int, default=0)
    return parser.parse_args()


def classify(path: Path) -> tuple[str, int] | None:
    for channel in ("depth", "seg", "rgb"):
        match = PASS_PATTERNS[channel].search(path.name)
        if match:
            return channel, int(match.group(1))
    return None


def read_exr(path: Path, channels: tuple[str, ...]) -> np.ndarray:
    exr = OpenEXR.InputFile(str(path))
    try:
        window = exr.header()["dataWindow"]
        width = window.max.x - window.min.x + 1
        height = window.max.y - window.min.y + 1
        pixel_type = Imath.PixelType(Imath.PixelType.FLOAT)
        arrays = [
            np.frombuffer(exr.channel(channel, pixel_type), dtype=np.float32).reshape(height, width)
            for channel in channels
        ]
        return arrays[0] if len(arrays) == 1 else np.stack(arrays, axis=-1)
    finally:
        exr.close()


def atomic_jpeg(path: Path, rgb: np.ndarray, quality: int) -> None:
    mapped = np.power(np.clip(rgb, 0.0, 1.0), 1.0 / 2.2)
    image = np.rint(mapped * 255.0).astype(np.uint8)
    path.parent.mkdir(parents=True, exist_ok=True)
    with tempfile.NamedTemporaryFile(dir=path.parent, suffix=".jpg", delete=False) as handle:
        temporary = Path(handle.name)
    try:
        imageio.imwrite(temporary, image, quality=quality)
        if temporary.stat().st_size == 0:
            raise RuntimeError("JPEG writer produced an empty file.")
        temporary.replace(path)
    finally:
        temporary.unlink(missing_ok=True)


def atomic_npz(path: Path, key: str, array: np.ndarray) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    with tempfile.NamedTemporaryFile(dir=path.parent, suffix=".npz", delete=False) as handle:
        temporary = Path(handle.name)
        np.savez_compressed(handle, **{key: array})
    try:
        with np.load(temporary) as data:
            if key not in data or data[key].shape != array.shape:
                raise RuntimeError(f"NPZ verification failed for {temporary}")
        temporary.replace(path)
    finally:
        temporary.unlink(missing_ok=True)


def convert(job: Job, jpeg_quality: int, delete_source: bool) -> tuple[str, Path]:
    if job.channel == "rgb":
        atomic_jpeg(job.destination, read_exr(job.source, ("R", "G", "B")), jpeg_quality)
    elif job.channel == "depth":
        atomic_npz(job.destination, "depth", read_exr(job.source, ("R",)))
    else:
        atomic_npz(job.destination, "seg", read_exr(job.source, ("R",)))
    if delete_source:
        job.source.unlink()
    return job.channel, job.destination


def copy_top_level_metadata(source: Path, destination: Path) -> None:
    if source.resolve() == destination.resolve():
        return
    destination.mkdir(parents=True, exist_ok=True)
    for path in source.iterdir():
        if path.is_file() and path.suffix.lower() in {".json", ".txt"}:
            shutil.copy2(path, destination / path.name)


def collect_jobs(source: Path, destination: Path, channels: set[str]) -> list[Job]:
    jobs: list[Job] = []
    for camera_dir in sorted(path for path in source.iterdir() if path.is_dir()):
        for exr_path in sorted(camera_dir.glob("*.exr")):
            identified = classify(exr_path)
            if identified is None:
                continue
            channel, frame = identified
            if channel not in channels:
                continue
            extension = ".jpg" if channel == "rgb" else ".npz"
            output = destination / camera_dir.name / channel / f"{frame:04d}{extension}"
            jobs.append(Job(exr_path, output, channel))
    return jobs


def validate_outputs(jobs: list[Job]) -> None:
    missing = [job.destination for job in jobs if not job.destination.is_file()]
    empty = [
        job.destination
        for job in jobs
        if job.destination.is_file() and job.destination.stat().st_size == 0
    ]
    if missing or empty:
        examples = [str(path) for path in (missing + empty)[:10]]
        raise RuntimeError(
            f"Output verification failed: missing={len(missing)}, empty={len(empty)}; "
            f"examples={examples}"
        )


def convert_directory(
    source: Path,
    destination: Path | None,
    channels: set[str],
    workers: int,
    jpeg_quality: int,
    delete_source: bool,
) -> dict[str, int]:
    source = source.resolve()
    if not source.is_dir():
        raise FileNotFoundError(f"Render directory does not exist: {source}")
    destination = source if destination is None else destination.resolve()
    destination.mkdir(parents=True, exist_ok=True)
    copy_top_level_metadata(source, destination)
    jobs = collect_jobs(source, destination, channels)
    if not jobs:
        raise RuntimeError(f"No recognized EXR passes found under {source}")

    counts = {channel: 0 for channel in sorted(channels)}
    errors: list[str] = []
    with ThreadPoolExecutor(max_workers=max(1, workers)) as executor:
        futures = {
            executor.submit(convert, job, jpeg_quality, delete_source): job for job in jobs
        }
        for future in tqdm(
            as_completed(futures), total=len(futures), desc=source.name, unit="file"
        ):
            job = futures[future]
            try:
                channel, _ = future.result()
                counts[channel] += 1
            except Exception as exc:
                errors.append(f"{job.source}: {exc}")

    if errors:
        raise RuntimeError(
            f"{len(errors)} conversion(s) failed. First errors:\n" + "\n".join(errors[:10])
        )
    validate_outputs(jobs)
    return counts


def main() -> int:
    args = parse_args()
    channels = {item.strip().lower() for item in args.channels.split(",") if item.strip()}
    unknown = channels - set(PASS_PATTERNS)
    if not channels or unknown:
        raise ValueError(f"Invalid --channels value; unknown={sorted(unknown)}")
    if not args.skip_ply and not {"rgb", "depth"}.issubset(channels):
        raise ValueError("PLY generation requires both rgb and depth; use --skip-ply otherwise.")
    if args.output_dir is not None and len(args.render_dirs) != 1:
        raise ValueError("--output-dir can only be used with one input directory.")

    failed = 0
    for source in args.render_dirs:
        try:
            counts = convert_directory(
                source,
                args.output_dir,
                channels,
                args.workers,
                args.jpeg_quality,
                args.delete_source_exr,
            )
            if not args.skip_ply:
                render_dir = source if args.output_dir is None else args.output_dir
                ply = generate_points3d(
                    render_dir,
                    depth_threshold=args.depth_threshold,
                    max_points=args.max_points,
                    seed=args.ply_seed,
                )
                counts["points3d"] = int(ply["points"])
            print(f"[SUCCESS] {source}: {json.dumps(counts, sort_keys=True)}")
        except Exception as exc:
            failed += 1
            print(f"[FAILED] {source}: {exc}")
    return 1 if failed else 0


if __name__ == "__main__":
    raise SystemExit(main())