simulation/datasets_construct.sh

#!/usr/bin/env bash
#
# Build and verify four-view images for dataset trajectories.
#
# For each trajectory CSV, this script runs run.sh in dataset playback mode.
# The Gazebo plugin stops at every waypoint and captures front/down/left/right.
# After every run, this script verifies that all images referenced by the CSV
# exist and rejects near-uniform frames or low-texture gray render patches.
# Failed trajectories are restored and retried.
#
# Usage:
#   ./datasets_construct.sh
#   ./datasets_construct.sh gui
#   ./datasets_construct.sh datasets/world_v3
#   DATASET_ROOT=./datasets/world_v3 ./datasets_construct.sh

set -euo pipefail

PROJECT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"

MODE_ARGS=()
if [[ "${1:-}" == "gui" ]]; then
  MODE_ARGS+=(gui)
  shift
fi

ROOTS=()
if [[ $# -gt 0 ]]; then
  for root in "$@"; do
    ROOTS+=("${root}")
  done
elif [[ -n "${DATASET_ROOT:-}" ]]; then
  ROOTS+=("${DATASET_ROOT}")
else
  ROOTS+=("${PROJECT_DIR}/datasets/world_v3")
fi

MAX_ATTEMPTS="${DATASET_CAPTURE_MAX_ATTEMPTS:-5}"
CAPTURE_LIMIT="${DATASET_CAPTURE_LIMIT:-}"
VALIDATOR_PY="$(mktemp /tmp/simple_drone_validate.XXXXXX.py)"
ORIGINAL_CSV=""
CURRENT_TRAJ=""

cleanup() {
  rm -f "${VALIDATOR_PY}"
  if [[ -n "${ORIGINAL_CSV}" && -f "${ORIGINAL_CSV}" && -n "${CURRENT_TRAJ}" ]]; then
    cp "${ORIGINAL_CSV}" "${CURRENT_TRAJ}" 2>/dev/null || true
  fi
  if [[ -n "${ORIGINAL_CSV}" && -f "${ORIGINAL_CSV}" ]]; then
    rm -f "${ORIGINAL_CSV}"
  fi
}
trap cleanup EXIT

cat > "${VALIDATOR_PY}" <<'PY'
import csv
import os
import shutil
import sys
from pathlib import Path

try:
    import numpy as np
    from PIL import Image
except Exception as exc:
    print(f"validator dependency error: {exc}", file=sys.stderr)
    raise SystemExit(2)

VIEWS = ("front", "down", "left", "right")
GRAY_DELTA = int(os.environ.get("DATASET_GRAY_DELTA", "3"))
GRAY_MIN = int(os.environ.get("DATASET_GRAY_MIN", "120"))
GRAY_MAX = int(os.environ.get("DATASET_GRAY_MAX", "136"))
GRAY_MAX_RATIO = min(0.99, max(0.0, float(os.environ.get("DATASET_GRAY_MAX_RATIO", "0.12"))))
GRAY_TILE_RATIO = min(0.99, max(0.0, float(os.environ.get("DATASET_GRAY_TILE_RATIO", "0.90"))))
GRAY_SCANLINE_RATIO = min(0.99, max(0.0, float(os.environ.get("DATASET_GRAY_SCANLINE_RATIO", "0.95"))))
MIN_LUMA_STD = float(os.environ.get("DATASET_MIN_LUMA_STD", "4.0"))
MIN_RGB_RANGE = int(os.environ.get("DATASET_MIN_RGB_RANGE", "12"))


def rows_and_required(csv_path: Path):
    with csv_path.open(newline="", encoding="utf-8") as f:
        reader = csv.DictReader(f)
        required = {
            "capture_frame",
            "captured_yaw_rad",
            "captured_yaw_deg",
            "front_image",
            "down_image",
            "left_image",
            "right_image",
        }
        if not reader.fieldnames or not required.issubset(reader.fieldnames):
            return None, "missing capture/image columns"
        rows = list(reader)
    if not rows:
        return None, "empty trajectory CSV"
    return rows, ""


def validate_image(path: Path):
    try:
        with Image.open(path) as image:
            rgb = image.convert("RGB")
            arr = np.asarray(rgb, dtype=np.uint8)
    except Exception as exc:
        return False, f"cannot open image: {path} ({exc})"

    if arr.ndim != 3 or arr.shape[0] < 8 or arr.shape[1] < 8:
        return False, f"bad image dimensions: {path}"

    channel_min = arr.min(axis=2).astype(np.int16)
    channel_max = arr.max(axis=2).astype(np.int16)
    luma = arr.mean(axis=2)
    rgb_range = int(arr.max()) - int(arr.min())
    luma_std = float(luma.std())

    if rgb_range < MIN_RGB_RANGE:
        return False, f"all/near-gray image: {path} range={rgb_range} luma_std={luma_std:.3f}"

    neutral = (channel_max - channel_min) <= GRAY_DELTA
    medium_gray = (luma >= GRAY_MIN) & (luma <= GRAY_MAX)
    gray_mask = neutral & medium_gray
    gray_ratio = float(gray_mask.mean())
    if gray_ratio > GRAY_MAX_RATIO:
        return False, f"global gray ratio too high: {path} ratio={gray_ratio:.4f}"

    buffer_mask = np.all(arr == 128, axis=2)
    row_ratios = buffer_mask.mean(axis=1)
    column_ratios = buffer_mask.mean(axis=0)
    if float(row_ratios.max()) > GRAY_SCANLINE_RATIO:
        row = int(row_ratios.argmax())
        return False, (
            f"gray render-buffer row: {path} row={row} "
            f"ratio={float(row_ratios[row]):.4f}"
        )
    if float(column_ratios.max()) > GRAY_SCANLINE_RATIO:
        column = int(column_ratios.argmax())
        return False, (
            f"gray render-buffer column: {path} column={column} "
            f"ratio={float(column_ratios[column]):.4f}"
        )

    height, width = gray_mask.shape
    tile_h = max(1, height // 8)
    tile_w = max(1, width // 8)
    for y0 in range(0, height, tile_h):
        for x0 in range(0, width, tile_w):
            tile = gray_mask[y0:min(y0 + tile_h, height), x0:min(x0 + tile_w, width)]
            tile_ratio = float(tile.mean()) if tile.size else 0.0
            tile_luma = luma[y0:min(y0 + tile_h, height), x0:min(x0 + tile_w, width)]
            tile_rgb = arr[y0:min(y0 + tile_h, height), x0:min(x0 + tile_w, width), :]
            tile_std = float(tile_luma.std()) if tile_luma.size else 0.0
            tile_range = int(tile_rgb.max()) - int(tile_rgb.min()) if tile_rgb.size else 0
            if tile_ratio > GRAY_TILE_RATIO and tile_std < MIN_LUMA_STD and tile_range < MIN_RGB_RANGE:
                return False, (
                    f"large low-texture gray tile: {path} "
                    f"tile_ratio={tile_ratio:.4f} range={tile_range} luma_std={tile_std:.3f}"
                )

    return True, ""


def validate(dataset_root: Path, csv_path: Path):
    rows, reason = rows_and_required(csv_path)
    if rows is None:
        return False, reason

    image_root = dataset_root / "images" / csv_path.stem
    for index, row in enumerate(rows):
        if not row.get("capture_frame") or not row.get("captured_yaw_rad") or not row.get("captured_yaw_deg"):
            return False, f"row {index} missing capture metadata"
        for view in VIEWS:
            image_name = row.get(f"{view}_image", "")
            if not image_name:
                return False, f"row {index} missing {view}_image"
            image_path = image_root / f"images_{view}" / image_name
            if not image_path.is_file():
                return False, f"missing image: {image_path}"
            ok, detail = validate_image(image_path)
            if not ok:
                return False, detail
    return True, "ok"


def quick_complete(dataset_root: Path, csv_path: Path):
    rows, reason = rows_and_required(csv_path)
    if rows is None:
        return False, reason

    image_root = dataset_root / "images" / csv_path.stem
    for index, row in enumerate(rows):
        if not row.get("capture_frame") or not row.get("captured_yaw_rad") or not row.get("captured_yaw_deg"):
            return False, f"row {index} missing capture metadata"
        for view in VIEWS:
            image_name = row.get(f"{view}_image", "")
            if not image_name:
                return False, f"row {index} missing {view}_image"
            image_path = image_root / f"images_{view}" / image_name
            if not image_path.is_file():
                return False, f"missing image: {image_path}"
    return True, "ok"


def clean(dataset_root: Path, csv_path: Path):
    image_root = dataset_root / "images" / csv_path.stem
    if image_root.exists():
        shutil.rmtree(image_root)


def main():
    if len(sys.argv) != 4:
        raise SystemExit("usage: validator.py validate|clean DATASET_ROOT CSV")
    command = sys.argv[1]
    dataset_root = Path(sys.argv[2])
    csv_path = Path(sys.argv[3])
    if command == "validate":
        ok, message = validate(dataset_root, csv_path)
        print(message)
        raise SystemExit(0 if ok else 1)
    if command == "quick":
        ok, message = quick_complete(dataset_root, csv_path)
        print(message)
        raise SystemExit(0 if ok else 1)
    if command == "clean":
        clean(dataset_root, csv_path)
        raise SystemExit(0)
    raise SystemExit(f"unknown command: {command}")


if __name__ == "__main__":
    main()
PY

trajectory_complete() {
  local dataset_root="$1"
  local csv_path="$2"
  if [[ "${DATASET_REVALIDATE_EXISTING:-0}" != "0" ]]; then
    python3 "${VALIDATOR_PY}" validate "${dataset_root}" "${csv_path}" >/dev/null 2>&1
  else
    python3 "${VALIDATOR_PY}" quick "${dataset_root}" "${csv_path}" >/dev/null 2>&1
  fi
}

validate_or_explain() {
  local dataset_root="$1"
  local csv_path="$2"
  python3 "${VALIDATOR_PY}" validate "${dataset_root}" "${csv_path}"
}

clean_outputs() {
  local dataset_root="$1"
  local csv_path="$2"
  python3 "${VALIDATOR_PY}" clean "${dataset_root}" "${csv_path}"
}

for root in "${ROOTS[@]}"; do
  DATASET_ROOT_ABS="$(cd "${root}" && pwd)"
  TRAJECTORY_DIR="${DATASET_ROOT_ABS}/trajectory"

  if [[ ! -d "${TRAJECTORY_DIR}" ]]; then
    echo "[datasets_construct] trajectory directory not found: ${TRAJECTORY_DIR}" >&2
    exit 1
  fi

  mapfile -t ALL_TRAJECTORIES < <(find "${TRAJECTORY_DIR}" -maxdepth 1 -type f -name '*.csv' | sort)
  if [[ -n "${DATASET_CAPTURE_TRAJECTORY:-}" ]]; then
    REQUESTED_TRAJECTORY="${DATASET_CAPTURE_TRAJECTORY}"
    if [[ "${REQUESTED_TRAJECTORY}" != /* ]]; then
      REQUESTED_TRAJECTORY="${TRAJECTORY_DIR}/${REQUESTED_TRAJECTORY}"
    fi
    if [[ ! -f "${REQUESTED_TRAJECTORY}" ]]; then
      echo "[datasets_construct] requested trajectory not found: ${REQUESTED_TRAJECTORY}" >&2
      exit 1
    fi
    ALL_TRAJECTORIES=("$(cd "$(dirname "${REQUESTED_TRAJECTORY}")" && pwd)/$(basename "${REQUESTED_TRAJECTORY}")")
  fi
  if [[ "${#ALL_TRAJECTORIES[@]}" -eq 0 ]]; then
    echo "[datasets_construct] no trajectory CSV files found in ${TRAJECTORY_DIR}" >&2
    exit 1
  fi

  echo "[datasets_construct] dataset -> ${DATASET_ROOT_ABS}"
  echo "[datasets_construct] total trajectories -> ${#ALL_TRAJECTORIES[@]}"
  echo "[datasets_construct] scanning existing captures..."

  TRAJECTORIES=()
  scanned=0
  for traj in "${ALL_TRAJECTORIES[@]}"; do
    scanned=$((scanned + 1))
    if (( scanned % 100 == 0 )); then
      echo "[datasets_construct] scanned ${scanned}/${#ALL_TRAJECTORIES[@]}"
    fi
    if trajectory_complete "${DATASET_ROOT_ABS}" "${traj}"; then
      echo "[datasets_construct] skip verified $(basename "${traj}")"
    else
      TRAJECTORIES+=("${traj}")
      if [[ -n "${CAPTURE_LIMIT}" && "${#TRAJECTORIES[@]}" -ge "${CAPTURE_LIMIT}" ]]; then
        echo "[datasets_construct] reached DATASET_CAPTURE_LIMIT=${CAPTURE_LIMIT}; stopping scan early"
        break
      fi
    fi
  done

  echo "[datasets_construct] remaining trajectories -> ${#TRAJECTORIES[@]}"
  echo "[datasets_construct] max attempts -> ${MAX_ATTEMPTS}"

  if [[ "${#TRAJECTORIES[@]}" -eq 0 ]]; then
    echo "[datasets_construct] complete: all trajectories are already verified"
    continue
  fi

  skipped_timeouts=0
  for i in "${!TRAJECTORIES[@]}"; do
    traj="${TRAJECTORIES[$i]}"
    CURRENT_TRAJ="${traj}"
    idx=$((i + 1))
    echo "[datasets_construct] ${idx}/${#TRAJECTORIES[@]} running $(basename "${traj}")"

    ORIGINAL_CSV="$(mktemp /tmp/simple_drone_csv.XXXXXX)"
    cp "${traj}" "${ORIGINAL_CSV}"

    attempt=1
    timed_out=0
    while [[ "${attempt}" -le "${MAX_ATTEMPTS}" ]]; do
      echo "[datasets_construct] attempt ${attempt}/${MAX_ATTEMPTS} $(basename "${traj}")"
      cp "${ORIGINAL_CSV}" "${traj}"
      clean_outputs "${DATASET_ROOT_ABS}" "${traj}"

      if DATASET_ROOT="${DATASET_ROOT_ABS}" SIMPLE_DRONE_QUIT_ON_FINISH=1 \
        SIMPLE_DRONE_RENDER_BACKEND="${SIMPLE_DRONE_RENDER_BACKEND:-system}" \
        SIMPLE_DRONE_SINGLE_CAMERA_SWEEP="${SIMPLE_DRONE_SINGLE_CAMERA_SWEEP:-0}" \
        "${PROJECT_DIR}/run.sh" "${MODE_ARGS[@]}" "${traj}"; then
        if validate_or_explain "${DATASET_ROOT_ABS}" "${traj}"; then
          echo "[datasets_construct] verified $(basename "${traj}")"
          rm -f "${ORIGINAL_CSV}"
          ORIGINAL_CSV=""
          CURRENT_TRAJ=""
          break
        fi
      else
        run_status=$?
        if [[ "${run_status}" -eq 124 ]]; then
          timed_out=1
          echo "[datasets_construct] timeout after ${SIMPLE_DRONE_RUN_TIMEOUT:-900}s; skipping $(basename "${traj}")" >&2
          break
        fi
        echo "[datasets_construct] run.sh failed for $(basename "${traj}") status=${run_status}" >&2
      fi

      attempt=$((attempt + 1))
      if [[ "${attempt}" -le "${MAX_ATTEMPTS}" ]]; then
        echo "[datasets_construct] retrying $(basename "${traj}")"
        sleep 1
      fi
    done

    if [[ -n "${ORIGINAL_CSV}" ]]; then
      cp "${ORIGINAL_CSV}" "${traj}"
      rm -f "${ORIGINAL_CSV}"
      ORIGINAL_CSV=""
      CURRENT_TRAJ=""
      if [[ "${timed_out}" -ne 0 ]]; then
        clean_outputs "${DATASET_ROOT_ABS}" "${traj}"
        skipped_timeouts=$((skipped_timeouts + 1))
        continue
      fi
      echo "[datasets_construct] failed after ${MAX_ATTEMPTS} attempts: ${traj}" >&2
      exit 1
    fi
  done
  if [[ "${skipped_timeouts}" -gt 0 ]]; then
    echo "[datasets_construct] skipped timed-out trajectories -> ${skipped_timeouts}" >&2
    exit 1
  fi
done

echo "[datasets_construct] complete"