"""Shared dataset discovery and image conversion for evaluation commands.""" from __future__ import annotations from dataclasses import dataclass from pathlib import Path import cv2 import numpy as np from lumenx_server.decode import DecodedImage, decode_upload IMAGE_SUFFIXES = {".jpg", ".jpeg", ".png", ".webp", ".tif", ".tiff", ".bmp"} @dataclass(frozen=True) class ImageRecord: relative_path: str input_path: Path reference_path: Path | None def _images(directory: Path) -> list[Path]: return sorted( (path for path in directory.rglob("*") if path.is_file() and path.suffix.lower() in IMAGE_SUFFIXES), key=lambda path: str(path).casefold(), ) def _discover_input_dir(root: Path) -> Path: preferred = ( "test/low", "eval/low", "test/low light", "real_captured/test/low", "synthetic/test/low", "low", ) by_lower = {str(path.relative_to(root)).replace("\\", "/").lower(): path for path in root.rglob("*") if path.is_dir()} for suffix in preferred: for relative, path in by_lower.items(): if relative.endswith(suffix): return path low_dirs = [path for path in by_lower.values() if path.name.lower() in {"low", "lowlight", "low_light"}] if len(low_dirs) == 1: return low_dirs[0] return root def _discover_reference_dir(input_dir: Path) -> Path | None: parent = input_dir.parent names = {path.name.lower(): path for path in parent.iterdir() if path.is_dir()} for name in ("high", "normal", "gt", "ground_truth", "reference", "target"): if name in names: return names[name] return None def discover_records( root: Path, *, input_dir: Path | None = None, reference_dir: Path | None = None, paired: bool = False, ) -> list[ImageRecord]: """Find low-light images and optionally match references by relative path/stem.""" root = root.expanduser().resolve() if not root.is_dir(): raise FileNotFoundError(f"Dataset root does not exist: {root}") low = (input_dir or _discover_input_dir(root)).expanduser().resolve() if not low.is_dir(): raise FileNotFoundError(f"Input image directory does not exist: {low}") high = reference_dir.expanduser().resolve() if reference_dir else _discover_reference_dir(low) if paired and (high is None or not high.is_dir()): raise FileNotFoundError( "Paired evaluation needs a reference directory. Pass --reference-dir or use sibling Low/High or Low/Normal folders." ) reference_by_stem: dict[str, Path] = {} if high is not None and high.is_dir(): for path in _images(high): reference_by_stem.setdefault(path.stem.casefold(), path) records: list[ImageRecord] = [] unmatched = 0 for path in _images(low): relative = str(path.relative_to(low)).replace("\\", "/") ref = None if high is not None: candidate = high / Path(relative) if candidate.is_file(): ref = candidate else: ref = reference_by_stem.get(path.stem.casefold()) if paired and ref is None: unmatched += 1 continue records.append(ImageRecord(relative, path, ref)) if paired and unmatched: raise FileNotFoundError(f"{unmatched} input image(s) under {low} have no matching reference under {high}") if not records: raise FileNotFoundError(f"No supported images found under {low}") if paired and not any(record.reference_path is not None for record in records): raise FileNotFoundError(f"No paired references matched images under {low}") return records def decode_path(path: Path, *, max_pixels: int = 12_000_000) -> DecodedImage: return decode_upload( path.read_bytes(), max_upload_bytes=max(path.stat().st_size, 1), max_pixels=max_pixels, hard_max_pixels=max(max_pixels, 50_000_000), ) def as_rgb_u8(decoded: DecodedImage) -> np.ndarray: return np.rint(np.clip(decoded.image, 0.0, 1.0) * 255.0).astype(np.uint8) def align_reference(reference: np.ndarray, shape: tuple[int, int]) -> np.ndarray: height, width = shape if reference.shape[:2] != (height, width): reference = cv2.resize(reference, (width, height), interpolation=cv2.INTER_AREA) return np.ascontiguousarray(reference.astype(np.uint8, copy=False))