dragon-ml-toolbox 12.13.0__py3-none-any.whl → 14.3.0__py3-none-any.whl

ml_tools/ML_vision_datasetmaster.py

@@ -0,0 +1,1352 @@
+import torch
+from torch.utils.data import Dataset, Subset
+import numpy
+from sklearn.model_selection import train_test_split
+from typing import Union, Tuple, List, Optional, Callable, Dict, Any
+from PIL import Image
+from torchvision.datasets import ImageFolder
+from torchvision import transforms
+import torchvision.transforms.functional as TF
+from pathlib import Path
+import random
+import json
+
+from .ML_datasetmaster import _BaseMaker
+from .path_manager import make_fullpath
+from ._logger import _LOGGER
+from ._script_info import _script_info
+from .keys import VisionTransformRecipeKeys, ObjectDetectionKeys
+from ._ML_vision_recipe import save_recipe
+from .ML_vision_transformers import TRANSFORM_REGISTRY
+from .custom_logger import custom_logger
+
+
+__all__ = [
+    "VisionDatasetMaker",
+    "SegmentationDatasetMaker",
+    "ObjectDetectionDatasetMaker"
+]
+
+
+# --- VisionDatasetMaker ---
+class VisionDatasetMaker(_BaseMaker):
+    """
+    Creates processed PyTorch datasets for computer vision tasks from an
+    image folder directory.
+    
+    Supports two modes:
+    1. `from_folder()`: Loads from one directory and splits into train/val/test.
+    2. `from_folders()`: Loads from pre-split train/val/test directories.
+    
+    Uses online augmentations per epoch (image augmentation without creating new files).
+    """
+    def __init__(self):
+        """
+        Typically not called directly. Use the class methods `from_folder()` or `from_folders()` to create an instance.
+        """
+        super().__init__()
+        self._full_dataset: Optional[ImageFolder] = None
+        self.labels: Optional[List[int]] = None
+        self.class_map: Optional[dict[str,int]] = None
+        
+        self._is_split = False
+        self._are_transforms_configured = False
+        self._val_recipe_components = None
+        self._has_mean_std: bool = False
+
+    @classmethod
+    def from_folder(cls, root_dir: Union[str,Path]) -> 'VisionDatasetMaker':
+        """
+        Creates a maker instance from a single root directory of images.
+        
+        This method assumes a single directory (e.g., 'data/') that
+        contains class subfolders (e.g., 'data/cat/', 'data/dog/').
+        
+        The dataset will be loaded in its entirety, and you MUST call
+        `.split_data()` afterward to create train/validation/test sets.
+
+        Args:
+            root_dir (str | Path): The path to the root directory containing class subfolders.
+
+        Returns:
+            VisionDatasetMaker: A new instance with the full dataset loaded.
+        """
+        root_path = make_fullpath(root_dir, enforce="directory")
+        # Load with NO transform. We get PIL Images.
+        full_dataset = ImageFolder(root=root_path, transform=None)
+        _LOGGER.info(f"Found {len(full_dataset)} images in {len(full_dataset.classes)} classes.")
+        
+        maker = cls()
+        maker._full_dataset = full_dataset
+        maker.labels = [s[1] for s in full_dataset.samples]
+        maker.class_map = full_dataset.class_to_idx
+        return maker
+    
+    @classmethod
+    def from_folders(cls, 
+                     train_dir: Union[str,Path], 
+                     val_dir: Union[str,Path], 
+                     test_dir: Optional[Union[str,Path]] = None) -> 'VisionDatasetMaker':
+        """
+        Creates a maker instance from separate, pre-split directories.
+        
+        This method is used when you already have 'train', 'val', and
+        optionally 'test' folders, each containing class subfolders.
+        It bypasses the need for `.split_data()`.
+
+        Args:
+            train_dir (str | Path): Path to the training data directory.
+            val_dir (str | Path): Path to the validation data directory.
+            test_dir (str | Path | None): Path to the test data directory.
+
+        Returns:
+            VisionDatasetMaker: A new, pre-split instance.
+
+        Raises:
+            ValueError: If the classes found in train, val, or test directories are inconsistent.
+        """
+        train_path = make_fullpath(train_dir, enforce="directory")
+        val_path = make_fullpath(val_dir, enforce="directory")
+        
+        _LOGGER.info("Loading data from separate directories.")
+        # Load with NO transform. We get PIL Images.
+        train_ds = ImageFolder(root=train_path, transform=None)
+        val_ds = ImageFolder(root=val_path, transform=None)
+        
+        # Check for class consistency
+        if train_ds.class_to_idx != val_ds.class_to_idx:
+            _LOGGER.error("Train and validation directories have different or inconsistent classes.")
+            raise ValueError()
+
+        maker = cls()
+        maker._train_dataset = train_ds
+        maker._val_dataset = val_ds
+        maker.class_map = train_ds.class_to_idx
+        
+        if test_dir:
+            test_path = make_fullpath(test_dir, enforce="directory")
+            test_ds = ImageFolder(root=test_path, transform=None)
+            if train_ds.class_to_idx != test_ds.class_to_idx:
+                _LOGGER.error("Train and test directories have different or inconsistent classes.")
+                raise ValueError()
+            maker._test_dataset = test_ds
+            _LOGGER.info(f"Loaded: {len(train_ds)} train, {len(val_ds)} val, {len(test_ds)} test images.")
+        else:
+            _LOGGER.info(f"Loaded: {len(train_ds)} train, {len(val_ds)} val images.")
+
+        maker._is_split = True # Mark as "split" since data is pre-split
+        return maker
+        
+    @staticmethod
+    def inspect_folder(path: Union[str, Path]):
+        """
+        Logs a report of the types, sizes, and channels of image files
+        found in the directory and its subdirectories.
+        
+        This is a utility method to help diagnose potential dataset
+        issues (e.g., mixed image modes, corrupted files) before loading.
+
+        Args:
+            path (str, Path): The directory path to inspect.
+        """
+        path_obj = make_fullpath(path)
+
+        non_image_files = set()
+        img_types = set()
+        img_sizes = set()
+        img_channels = set()
+        img_counter = 0
+
+        _LOGGER.info(f"Inspecting folder: {path_obj}...")
+        # Use rglob to recursively find all files
+        for filepath in path_obj.rglob('*'):
+            if filepath.is_file():
+                try:
+                    # Using PIL to open is a more reliable check
+                    with Image.open(filepath) as img:
+                        img_types.add(img.format)
+                        img_sizes.add(img.size)
+                        img_channels.update(img.getbands())
+                        img_counter += 1
+                except (IOError, SyntaxError):
+                    non_image_files.add(filepath.name)
+
+        if non_image_files:
+            _LOGGER.warning(f"Non-image or corrupted files found and ignored: {non_image_files}")
+
+        report = (
+            f"\n--- Inspection Report for '{path_obj.name}' ---\n"
+            f"Total images found: {img_counter}\n"
+            f"Image formats: {img_types or 'None'}\n"
+            f"Image sizes (WxH): {img_sizes or 'None'}\n"
+            f"Image channels (bands): {img_channels or 'None'}\n"
+            f"--------------------------------------"
+        )
+        print(report)
+
+    def split_data(self, val_size: float = 0.2, test_size: float = 0.0, 
+                   stratify: bool = True, random_state: Optional[int] = None) -> 'VisionDatasetMaker':
+        """
+        Splits the dataset into train, validation, and optional test sets.
+        
+        This method MUST be called if you used `from_folder()`. It has no effect if you used `from_folders()`.
+
+        Args:
+            val_size (float): Proportion of the dataset to reserve for
+                              validation (e.g., 0.2 for 20%).
+            test_size (float): Proportion of the dataset to reserve for
+                               testing.
+            stratify (bool): If True, splits are performed in a stratified
+                             fashion, preserving the class distribution.
+            random_state (int | None): Seed for the random number generator for reproducible splits.
+
+        Returns:
+            VisionDatasetMaker: The same instance, now with datasets split.
+            
+        Raises:
+            ValueError: If `val_size` and `test_size` sum to 1.0 or more.
+        """
+        if self._is_split:
+            _LOGGER.warning("Data has already been split.")
+            return self
+
+        if val_size + test_size >= 1.0:
+            _LOGGER.error("The sum of val_size and test_size must be less than 1.")
+            raise ValueError()
+        
+        if not self._full_dataset:
+            _LOGGER.error("There is no dataset to split.")
+            raise ValueError()
+        
+        indices = list(range(len(self._full_dataset)))
+        labels_for_split = self.labels if stratify else None
+
+        train_indices, val_test_indices = train_test_split(
+            indices, test_size=(val_size + test_size), random_state=random_state, stratify=labels_for_split
+        )
+        
+        if not self.labels:
+            _LOGGER.error("Error when getting full dataset labels.")
+            raise ValueError()
+
+        if test_size > 0:
+            val_test_labels = [self.labels[i] for i in val_test_indices]
+            stratify_val_test = val_test_labels if stratify else None
+            val_indices, test_indices = train_test_split(
+                val_test_indices, test_size=(test_size / (val_size + test_size)), 
+                random_state=random_state, stratify=stratify_val_test
+            )
+            self._test_dataset = Subset(self._full_dataset, test_indices)
+            _LOGGER.info(f"Test set created with {len(self._test_dataset)} images.")
+        else:
+            val_indices = val_test_indices
+        
+        self._train_dataset = Subset(self._full_dataset, train_indices)
+        self._val_dataset = Subset(self._full_dataset, val_indices)
+        self._is_split = True
+        
+        _LOGGER.info(f"Data split into: \n- Training: {len(self._train_dataset)} images \n- Validation: {len(self._val_dataset)} images")
+        return self
+
+    def configure_transforms(self, resize_size: int = 256, crop_size: int = 224, 
+                             mean: Optional[List[float]] = [0.485, 0.456, 0.406], 
+                             std: Optional[List[float]] = [0.229, 0.224, 0.225],
+                             pre_transforms: Optional[List[Callable]] = None,
+                             extra_train_transforms: Optional[List[Callable]] = None) -> 'VisionDatasetMaker':
+        """
+        Configures and applies the image transformations and augmentations.
+        
+        This method must be called AFTER data is loaded and split.
+        
+        It sets up two pipelines:
+        1.  **Training Pipeline:** Includes random augmentations like
+            `RandomResizedCrop` and `RandomHorizontalFlip` (plus any
+            `extra_train_transforms`) for online augmentation.
+        2.  **Validation/Test Pipeline:** A deterministic pipeline using
+            `Resize` and `CenterCrop` for consistent evaluation.
+            
+        Both pipelines finish with `ToTensor` and `Normalize`.
+
+        Args:
+            resize_size (int): The size to resize the smallest edge to
+                               for validation/testing.
+            crop_size (int): The target size (square) for the final
+                             cropped image.
+            mean (List[float]): The mean values for normalization (e.g., ImageNet mean).
+            std (List[float]): The standard deviation values for normalization (e.g., ImageNet std).
+            extra_train_transforms (List[Callable] | None): A list of additional torchvision transforms to add to the end of the training transformations.
+            pre_transforms (List[Callable] | None): An list of transforms to be applied at the very beginning of the transformations for all sets.
+
+        Returns:
+            VisionDatasetMaker: The same instance, with transforms applied.
+            
+        Raises:
+            RuntimeError: If called before data is split.
+        """
+        if not self._is_split:
+            _LOGGER.error("Transforms must be configured AFTER splitting data (or using `from_folders`). Call .split_data() first if using `from_folder`.")
+            raise RuntimeError()
+        
+        if (mean is None and std is not None) or (mean is not None and std is None):
+            _LOGGER.error(f"'mean' and 'std' must be both None or both defined, but only one was provided.")
+            raise ValueError()
+
+        # --- Define Transform Pipelines ---
+        # These now MUST include ToTensor and Normalize, as the ImageFolder was loaded with transform=None.
+        
+        # --- Store components for validation recipe ---
+        self._val_recipe_components = {
+            VisionTransformRecipeKeys.PRE_TRANSFORMS: pre_transforms or [],
+            VisionTransformRecipeKeys.RESIZE_SIZE: resize_size,
+            VisionTransformRecipeKeys.CROP_SIZE: crop_size,
+        }
+        
+        if mean is not None and std is not None:
+            self._val_recipe_components.update({
+                VisionTransformRecipeKeys.MEAN: mean,
+                VisionTransformRecipeKeys.STD: std
+            })
+            self._has_mean_std = True
+        
+        base_pipeline = []
+        if pre_transforms:
+            base_pipeline.extend(pre_transforms)
+
+        # Base augmentations for training
+        base_train_transforms = [
+            transforms.RandomResizedCrop(crop_size), 
+            transforms.RandomHorizontalFlip()
+        ]
+        if extra_train_transforms:
+            base_train_transforms.extend(extra_train_transforms)
+        
+        # Final conversion and normalization
+        final_transforms: list[Callable] = [
+            transforms.ToTensor()
+        ]
+        
+        if self._has_mean_std:
+            final_transforms.append(transforms.Normalize(mean=mean, std=std))
+
+        # Build the val/test pipeline
+        val_transform_list = [
+            *base_pipeline, # Apply pre_transforms first
+            transforms.Resize(resize_size), 
+            transforms.CenterCrop(crop_size), 
+            *final_transforms
+        ]
+        
+        # Build the train pipeline
+        train_transform_list = [
+            *base_pipeline, # Apply pre_transforms first
+            *base_train_transforms, 
+            *final_transforms
+        ]
+        
+        val_transform = transforms.Compose(val_transform_list)
+        train_transform = transforms.Compose(train_transform_list)
+
+        # --- Apply Transforms using the Wrapper ---
+        # This correctly assigns the transform regardless of whether the dataset is a Subset (from_folder) or an ImageFolder (from_folders).
+        
+        self._train_dataset = _DatasetTransformer(self._train_dataset, train_transform) # type: ignore
+        self._val_dataset = _DatasetTransformer(self._val_dataset, val_transform) # type: ignore
+        if self._test_dataset:
+            self._test_dataset = _DatasetTransformer(self._test_dataset, val_transform) # type: ignore
+        
+        self._are_transforms_configured = True
+        _LOGGER.info("Image transforms configured and applied.")
+        return self
+
+    def get_datasets(self) -> Tuple[Dataset, ...]:
+        """
+        Returns the final train, validation, and optional test datasets.
+        
+        This is the final step, used to retrieve the datasets for use in
+        a `MLTrainer` or `DataLoader`.
+
+        Returns:
+            (Tuple[Dataset, ...]): A tuple containing the (train, val)
+                                 or (train, val, test) datasets.
+                                 
+        Raises:
+            RuntimeError: If called before data is split.
+            UserWarning: If called before transforms are configured.
+        """
+        if not self._is_split:
+            _LOGGER.error("Data has not been split. Call .split_data() first.")
+            raise RuntimeError()
+        if not self._are_transforms_configured:
+            _LOGGER.warning("Transforms have not been configured.")
+
+        if self._test_dataset:
+            return self._train_dataset, self._val_dataset, self._test_dataset
+        return self._train_dataset, self._val_dataset
+
+    def save_transform_recipe(self, filepath: Union[str, Path]) -> None:
+        """
+        Saves the validation transform pipeline as a JSON recipe file.
+        
+        This recipe can be loaded by the PyTorchVisionInferenceHandler
+        to ensure identical preprocessing.
+
+        Args:
+            filepath (str | Path): The path to save the .json recipe file.
+        """
+        if not self._are_transforms_configured:
+            _LOGGER.error("Transforms are not configured. Call .configure_transforms() first.")
+            raise RuntimeError()
+
+        recipe: Dict[str, Any] = {
+            VisionTransformRecipeKeys.TASK: "classification",
+            VisionTransformRecipeKeys.PIPELINE: []
+        }
+        
+        components = self._val_recipe_components
+        
+        if not components:
+            _LOGGER.error(f"Error getting the transformers recipe for validation set.")
+            raise ValueError()
+        
+        # validate path
+        file_path = make_fullpath(filepath, make=True, enforce="file")
+
+        # 1. Handle pre_transforms
+        for t in components[VisionTransformRecipeKeys.PRE_TRANSFORMS]:
+            t_name = t.__class__.__name__
+            if t_name in TRANSFORM_REGISTRY:
+                recipe[VisionTransformRecipeKeys.PIPELINE].append({
+                    VisionTransformRecipeKeys.NAME: t_name,
+                    VisionTransformRecipeKeys.KWARGS: getattr(t, VisionTransformRecipeKeys.KWARGS, {})
+                })
+            else:
+                _LOGGER.warning(f"Skipping unknown pre_transform '{t_name}' in recipe. Not in TRANSFORM_REGISTRY.")
+                
+        # 2. Add standard transforms
+        recipe[VisionTransformRecipeKeys.PIPELINE].extend([
+            {VisionTransformRecipeKeys.NAME: "Resize", "kwargs": {"size": components[VisionTransformRecipeKeys.RESIZE_SIZE]}},
+            {VisionTransformRecipeKeys.NAME: "CenterCrop", "kwargs": {"size": components[VisionTransformRecipeKeys.CROP_SIZE]}},
+            {VisionTransformRecipeKeys.NAME: "ToTensor", "kwargs": {}}
+        ])
+        
+        if self._has_mean_std:
+            recipe[VisionTransformRecipeKeys.PIPELINE].append(
+                {VisionTransformRecipeKeys.NAME: "Normalize", "kwargs": {
+                "mean": components[VisionTransformRecipeKeys.MEAN],
+                "std": components[VisionTransformRecipeKeys.STD]
+                }}
+            )
+        
+        # 3. Save the file
+        save_recipe(recipe, file_path)
+        
+    def save_class_map(self, save_dir: Union[str,Path]) -> dict[str,int]:
+        """
+        Saves the class to index mapping {str: int} to a directory.
+        """
+        if not self.class_map:
+            _LOGGER.error(f"Class to index mapping is empty.")
+            raise ValueError()
+        
+        custom_logger(data=self.class_map,
+                      save_directory=save_dir,
+                      log_name="Class_to_Index",
+                      add_timestamp=False,
+                      dict_as="json")
+        
+        return self.class_map
+    
+
+class _DatasetTransformer(Dataset):
+    """
+    Internal wrapper class to apply a specific transform pipeline to any
+    dataset (e.g., a full ImageFolder or a Subset).
+    """
+    def __init__(self, dataset: Dataset, transform: Optional[transforms.Compose] = None):
+        self.dataset = dataset
+        self.transform = transform
+        
+        # --- Propagate attributes for inspection ---
+        # For ImageFolder
+        if hasattr(dataset, 'class_to_idx'):
+            self.class_to_idx = getattr(dataset, 'class_to_idx')
+        if hasattr(dataset, 'classes'):
+            self.classes = getattr(dataset, 'classes')
+        # For Subset
+        if hasattr(dataset, 'indices'):
+            self.indices = getattr(dataset, 'indices')
+        if hasattr(dataset, 'dataset'):
+            # This allows access to the *original* full dataset
+            self.original_dataset = getattr(dataset, 'dataset')
+
+    def __getitem__(self, index):
+        # Get the original data (e.g., PIL Image, label)
+        x, y = self.dataset[index] 
+        
+        # Apply the specific transform for this dataset
+        if self.transform:
+            x = self.transform(x)
+        return x, y
+        
+    def __len__(self):
+        return len(self.dataset) # type: ignore
+
+
+# --- Segmentation dataset ----
+class _SegmentationDataset(Dataset):
+    """
+    Internal helper class to load image-mask pairs.
+    
+    Loads images as RGB and masks as 'L' (grayscale, 8-bit integer pixels).
+    """
+    def __init__(self, image_paths: List[Path], mask_paths: List[Path], transform: Optional[Callable] = None):
+        self.image_paths = image_paths
+        self.mask_paths = mask_paths
+        self.transform = transform
+        
+        # --- Propagate 'classes' if they exist (for MLTrainer) ---
+        self.classes: List[str] = []
+
+    def __len__(self):
+        return len(self.image_paths)
+
+    def __getitem__(self, idx):
+        img_path = self.image_paths[idx]
+        mask_path = self.mask_paths[idx]
+        
+        try:
+            # Open as PIL Images. Masks should be 'L'
+            image = Image.open(img_path).convert("RGB")
+            mask = Image.open(mask_path).convert("L")
+        except Exception as e:
+            _LOGGER.error(f"Error loading sample #{idx}: {img_path.name} / {mask_path.name}. Error: {e}")
+            # Return empty tensors
+            return torch.empty(3, 224, 224), torch.empty(224, 224, dtype=torch.long)
+            
+        if self.transform:
+            image, mask = self.transform(image, mask)
+            
+        return image, mask
+
+
+# Internal Paired Transform Helpers
+class _PairedCompose:
+    """A 'Compose' for paired image/mask transforms."""
+    def __init__(self, transforms: List[Callable]):
+        self.transforms = transforms
+
+    def __call__(self, image: Any, mask: Any) -> Tuple[Any, Any]:
+        for t in self.transforms:
+            image, mask = t(image, mask)
+        return image, mask
+
+class _PairedToTensor:
+    """Converts a PIL Image pair (image, mask) to Tensors."""
+    def __call__(self, image: Image.Image, mask: Image.Image) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Use new variable names to satisfy the linter
+        image_tensor = TF.to_tensor(image)
+        # Convert mask to LongTensor, not float.
+        # This creates a [H, W] tensor of integer class IDs.
+        mask_tensor = torch.from_numpy(numpy.array(mask, dtype=numpy.int64))
+        return image_tensor, mask_tensor
+
+class _PairedNormalize:
+    """Normalizes the image tensor and leaves the mask untouched."""
+    def __init__(self, mean: List[float], std: List[float]):
+        self.normalize = transforms.Normalize(mean, std)
+    
+    def __call__(self, image: torch.Tensor, mask: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        image = self.normalize(image)
+        return image, mask
+
+class _PairedResize:
+    """Resizes an image and mask to the same size."""
+    def __init__(self, size: int):
+        self.size = [size, size]
+    
+    def __call__(self, image: Image.Image, mask: Image.Image) -> Tuple[Image.Image, Image.Image]:
+        # Use new variable names to avoid linter confusion
+        resized_image = TF.resize(image, self.size, interpolation=TF.InterpolationMode.BILINEAR) # type: ignore
+        # Use NEAREST for mask to avoid interpolating class IDs (e.g., 1.5)
+        resized_mask = TF.resize(mask, self.size, interpolation=TF.InterpolationMode.NEAREST) # type: ignore
+        return resized_image, resized_mask # type: ignore
+        
+class _PairedCenterCrop:
+    """Center-crops an image and mask to the same size."""
+    def __init__(self, size: int):
+        self.size = [size, size]
+    
+    def __call__(self, image: Image.Image, mask: Image.Image) -> Tuple[Image.Image, Image.Image]:
+        cropped_image = TF.center_crop(image, self.size) # type: ignore
+        cropped_mask = TF.center_crop(mask, self.size) # type: ignore
+        return cropped_image, cropped_mask # type: ignore
+
+class _PairedRandomHorizontalFlip:
+    """Applies the same random horizontal flip to both image and mask."""
+    def __init__(self, p: float = 0.5):
+        self.p = p
+    
+    def __call__(self, image: Image.Image, mask: Image.Image) -> Tuple[Image.Image, Image.Image]:
+        if random.random() < self.p:
+            flipped_image = TF.hflip(image) # type: ignore
+            flipped_mask = TF.hflip(mask) # type: ignore
+        return flipped_image, flipped_mask # type: ignore
+        
+class _PairedRandomResizedCrop:
+    """Applies the same random resized crop to both image and mask."""
+    def __init__(self, size: int, scale: Tuple[float, float]=(0.08, 1.0), ratio: Tuple[float, float]=(3./4., 4./3.)):
+        self.size = [size, size]
+        self.scale = scale
+        self.ratio = ratio
+        self.interpolation = TF.InterpolationMode.BILINEAR
+        self.mask_interpolation = TF.InterpolationMode.NEAREST
+
+    def __call__(self, image: Image.Image, mask: Image.Image) -> Tuple[Image.Image, Image.Image]:
+        # Get parameters for the random crop
+        # Convert scale/ratio tuples to lists to satisfy the linter's type hint
+        i, j, h, w = transforms.RandomResizedCrop.get_params(image, list(self.scale), list(self.ratio)) # type: ignore
+        
+        # Apply the crop with the SAME parameters and use new variable names
+        cropped_image = TF.resized_crop(image, i, j, h, w, self.size, self.interpolation) # type: ignore
+        cropped_mask = TF.resized_crop(mask, i, j, h, w, self.size, self.mask_interpolation) # type: ignore
+        
+        return cropped_image, cropped_mask # type: ignore
+
+# --- SegmentationDatasetMaker ---
+class SegmentationDatasetMaker(_BaseMaker):
+    """
+    Creates processed PyTorch datasets for segmentation from image and mask folders.
+
+    This maker finds all matching image-mask pairs from two directories,
+    splits them, and applies identical transformations (including augmentations)
+    to both the image and its corresponding mask.
+    
+    Workflow:
+    1. `maker = SegmentationDatasetMaker.from_folders(img_dir, mask_dir)`
+    2. `maker.set_class_map({'background': 0, 'road': 1})`
+    3. `maker.split_data(val_size=0.2)`
+    4. `maker.configure_transforms(crop_size=256)`
+    5. `train_ds, val_ds = maker.get_datasets()`
+    """
+    IMG_EXTENSIONS = ('.jpg', '.jpeg', '.png', '.bmp', '.tif', '.tiff')
+    
+    def __init__(self):
+        """
+        Typically not called directly. Use the class method `from_folders()` to create an instance.
+        """
+        super().__init__()
+        self.image_paths: List[Path] = []
+        self.mask_paths: List[Path] = []
+        self.class_map: Dict[str, int] = {}
+        
+        self._is_split = False
+        self._are_transforms_configured = False
+        self.train_transform: Optional[Callable] = None
+        self.val_transform: Optional[Callable] = None
+
+    @classmethod
+    def from_folders(cls, image_dir: Union[str, Path], mask_dir: Union[str, Path]) -> 'SegmentationDatasetMaker':
+        """
+        Creates a maker instance by loading all matching image-mask pairs
+        from two corresponding directories.
+        
+        This method assumes that for an image `images/img_001.png`, there
+        is a corresponding mask `masks/img_001.png`.
+        
+        Args:
+            image_dir (str | Path): Path to the directory containing input images.
+            mask_dir (str | Path): Path to the directory containing segmentation masks.
+
+        Returns:
+            SegmentationDatasetMaker: A new instance with all pairs loaded.
+        """
+        maker = cls()
+        img_path_obj = make_fullpath(image_dir, enforce="directory")
+        msk_path_obj = make_fullpath(mask_dir, enforce="directory")
+
+        # Find all images
+        image_files = sorted([
+            p for p in img_path_obj.glob('*.*') 
+            if p.suffix.lower() in cls.IMG_EXTENSIONS
+        ])
+        
+        if not image_files:
+            _LOGGER.error(f"No images with extensions {cls.IMG_EXTENSIONS} found in {image_dir}")
+            raise FileNotFoundError()
+
+        _LOGGER.info(f"Found {len(image_files)} images. Searching for matching masks in {mask_dir}...")
+        
+        good_img_paths = []
+        good_mask_paths = []
+
+        for img_file in image_files:
+            mask_file = None
+            
+            # 1. Try to find mask with the exact same name
+            mask_file_primary = msk_path_obj / img_file.name
+            if mask_file_primary.exists():
+                mask_file = mask_file_primary
+            
+            # 2. If not, try to find mask with same stem + common mask extension
+            if mask_file is None:
+                for ext in cls.IMG_EXTENSIONS: # Masks are often .png
+                    mask_file_secondary = msk_path_obj / (img_file.stem + ext)
+                    if mask_file_secondary.exists():
+                        mask_file = mask_file_secondary
+                        break
+            
+            # 3. If a match is found, add the pair
+            if mask_file:
+                good_img_paths.append(img_file)
+                good_mask_paths.append(mask_file)
+            else:
+                _LOGGER.warning(f"No corresponding mask found for image: {img_file.name}")
+        
+        if not good_img_paths:
+            _LOGGER.error("No matching image-mask pairs were found.")
+            raise FileNotFoundError()
+            
+        _LOGGER.info(f"Successfully found {len(good_img_paths)} image-mask pairs.")
+        maker.image_paths = good_img_paths
+        maker.mask_paths = good_mask_paths
+        
+        return maker
+
+    @staticmethod
+    def inspect_folder(path: Union[str, Path]):
+        """
+        Logs a report of the types, sizes, and channels of image files
+        found in the directory. Useful for checking masks.
+        """
+        VisionDatasetMaker.inspect_folder(path)
+
+    def set_class_map(self, class_map: Dict[str, int]) -> 'SegmentationDatasetMaker':
+        """
+        Sets a map of pixel_value -> class_name. This is used by the MLTrainer for clear evaluation reports.
+
+        Args:
+            class_map (Dict[int, str]): A dictionary mapping the integer pixel
+                value in a mask to its string name.
+                Example: {'background': 0, 'road': 1, 'car': 2}
+        """
+        self.class_map = class_map
+        _LOGGER.info(f"Class map set: {class_map}")
+        return self
+    
+    @property
+    def classes(self) -> List[str]:
+        """Returns the list of class names, if set."""
+        if self.class_map:
+            return list(self.class_map.keys())
+        return []
+
+    def split_data(self, val_size: float = 0.2, test_size: float = 0.0, 
+                   random_state: Optional[int] = 42) -> 'SegmentationDatasetMaker':
+        """
+        Splits the loaded image-mask pairs into train, validation, and test sets.
+
+        Args:
+            val_size (float): Proportion of the dataset to reserve for validation.
+            test_size (float): Proportion of the dataset to reserve for testing.
+            random_state (int | None): Seed for reproducible splits.
+
+        Returns:
+            SegmentationDatasetMaker: The same instance, now with datasets created.
+        """
+        if self._is_split:
+            _LOGGER.warning("Data has already been split.")
+            return self
+
+        if val_size + test_size >= 1.0:
+            _LOGGER.error("The sum of val_size and test_size must be less than 1.")
+            raise ValueError()
+        
+        if not self.image_paths:
+            _LOGGER.error("There is no data to split. Use .from_folders() first.")
+            raise RuntimeError()
+        
+        indices = list(range(len(self.image_paths)))
+
+        # Split indices
+        train_indices, val_test_indices = train_test_split(
+            indices, test_size=(val_size + test_size), random_state=random_state
+        )
+        
+        # Helper to get paths from indices
+        def get_paths(idx_list):
+            return [self.image_paths[i] for i in idx_list], [self.mask_paths[i] for i in idx_list]
+
+        train_imgs, train_masks = get_paths(train_indices)
+        
+        if test_size > 0:
+            val_indices, test_indices = train_test_split(
+                val_test_indices, test_size=(test_size / (val_size + test_size)), 
+                random_state=random_state
+            )
+            val_imgs, val_masks = get_paths(val_indices)
+            test_imgs, test_masks = get_paths(test_indices)
+            
+            self._test_dataset = _SegmentationDataset(test_imgs, test_masks, transform=None)
+            self._test_dataset.classes = self.classes # type: ignore
+            _LOGGER.info(f"Test set created with {len(self._test_dataset)} images.")
+        else:
+            val_imgs, val_masks = get_paths(val_test_indices)
+        
+        self._train_dataset = _SegmentationDataset(train_imgs, train_masks, transform=None)
+        self._val_dataset = _SegmentationDataset(val_imgs, val_masks, transform=None)
+        
+        # Propagate class names to datasets for MLTrainer
+        self._train_dataset.classes = self.classes # type: ignore
+        self._val_dataset.classes = self.classes # type: ignore
+
+        self._is_split = True
+        _LOGGER.info(f"Data split into: \n- Training: {len(self._train_dataset)} images \n- Validation: {len(self._val_dataset)} images")
+        return self
+
+    def configure_transforms(self, 
+                             resize_size: int = 256, 
+                             crop_size: int = 224, 
+                             mean: List[float] = [0.485, 0.456, 0.406], 
+                             std: List[float] = [0.229, 0.224, 0.225]) -> 'SegmentationDatasetMaker':
+        """
+        Configures and applies the image and mask transformations.
+        
+        This method must be called AFTER data is split.
+
+        Args:
+            resize_size (int): The size to resize the smallest edge to
+                               for validation/testing.
+            crop_size (int): The target size (square) for the final
+                             cropped image.
+            mean (List[float]): The mean values for image normalization.
+            std (List[float]): The std dev values for image normalization.
+
+        Returns:
+            SegmentationDatasetMaker: The same instance, with transforms applied.
+        """
+        if not self._is_split:
+            _LOGGER.error("Transforms must be configured AFTER splitting data. Call .split_data() first.")
+            raise RuntimeError()
+        
+        # --- Store components for validation recipe ---
+        self.val_recipe_components = {
+            VisionTransformRecipeKeys.RESIZE_SIZE: resize_size,
+            VisionTransformRecipeKeys.CROP_SIZE: crop_size,
+            VisionTransformRecipeKeys.MEAN: mean,
+            VisionTransformRecipeKeys.STD: std
+        }
+
+        # --- Validation/Test Pipeline (Deterministic) ---
+        self.val_transform = _PairedCompose([
+            _PairedResize(resize_size),
+            _PairedCenterCrop(crop_size),
+            _PairedToTensor(),
+            _PairedNormalize(mean, std)
+        ])
+        
+        # --- Training Pipeline (Augmentation) ---
+        self.train_transform = _PairedCompose([
+            _PairedRandomResizedCrop(crop_size),
+            _PairedRandomHorizontalFlip(p=0.5),
+            _PairedToTensor(),
+            _PairedNormalize(mean, std)
+        ])
+
+        # --- Apply Transforms to the Datasets ---
+        self._train_dataset.transform = self.train_transform # type: ignore
+        self._val_dataset.transform = self.val_transform # type: ignore
+        if self._test_dataset:
+            self._test_dataset.transform = self.val_transform # type: ignore
+        
+        self._are_transforms_configured = True
+        _LOGGER.info("Paired segmentation transforms configured and applied.")
+        return self
+
+    def get_datasets(self) -> Tuple[Dataset, ...]:
+        """
+        Returns the final train, validation, and optional test datasets.
+        
+        Raises:
+            RuntimeError: If called before data is split.
+            RuntimeError: If called before transforms are configured.
+        """
+        if not self._is_split:
+            _LOGGER.error("Data has not been split. Call .split_data() first.")
+            raise RuntimeError()
+        if not self._are_transforms_configured:
+            _LOGGER.error("Transforms have not been configured. Call .configure_transforms() first.")
+            raise RuntimeError()
+
+        if self._test_dataset:
+            return self._train_dataset, self._val_dataset, self._test_dataset
+        return self._train_dataset, self._val_dataset
+    
+    def save_transform_recipe(self, filepath: Union[str, Path]) -> None:
+        """
+        Saves the validation transform pipeline as a JSON recipe file.
+        
+        This recipe can be loaded by the PyTorchVisionInferenceHandler
+        to ensure identical preprocessing.
+
+        Args:
+            filepath (str | Path): The path to save the .json recipe file.
+        """
+        if not self._are_transforms_configured:
+            _LOGGER.error("Transforms are not configured. Call .configure_transforms() first.")
+            raise RuntimeError()
+        
+        components = self.val_recipe_components
+        
+        if not components:
+            _LOGGER.error(f"Error getting the transformers recipe for validation set.")
+            raise ValueError()
+        
+        # validate path
+        file_path = make_fullpath(filepath, make=True, enforce="file")
+
+        # Add standard transforms
+        recipe: Dict[str, Any] = {
+            VisionTransformRecipeKeys.TASK: "segmentation",
+            VisionTransformRecipeKeys.PIPELINE: [
+                {VisionTransformRecipeKeys.NAME: "Resize", "kwargs": {"size": components["resize_size"]}},
+                {VisionTransformRecipeKeys.NAME: "CenterCrop", "kwargs": {"size": components["crop_size"]}},
+                {VisionTransformRecipeKeys.NAME: "ToTensor", "kwargs": {}},
+                {VisionTransformRecipeKeys.NAME: "Normalize", "kwargs": {
+                    "mean": components["mean"],
+                    "std": components["std"]
+                }}
+            ]
+        }
+        
+        # Save the file
+        save_recipe(recipe, file_path)
+
+
+# Object detection
+def _od_collate_fn(batch: List[Tuple[torch.Tensor, Dict[str, torch.Tensor]]]) -> Tuple[List[torch.Tensor], List[Dict[str, torch.Tensor]]]:
+    """
+    Custom collate function for object detection.
+    
+    Takes a list of (image, target) tuples and zips them into two lists:
+    (list_of_images, list_of_targets).
+    This is required for models like Faster R-CNN, which accept a list
+    of images of varying sizes.
+    """
+    return tuple(zip(*batch)) # type: ignore
+
+
+class _ObjectDetectionDataset(Dataset):
+    """
+    Internal helper class to load image-annotation pairs.
+    
+    Loads an image as 'RGB' and parses its corresponding JSON annotation file
+    to create the required target dictionary (boxes, labels).
+    """
+    def __init__(self, image_paths: List[Path], annotation_paths: List[Path], transform: Optional[Callable] = None):
+        self.image_paths = image_paths
+        self.annotation_paths = annotation_paths
+        self.transform = transform
+        
+        # --- Propagate 'classes' if they exist (for MLTrainer) ---
+        self.classes: List[str] = []
+
+    def __len__(self):
+        return len(self.image_paths)
+
+    def __getitem__(self, idx):
+        img_path = self.image_paths[idx]
+        ann_path = self.annotation_paths[idx]
+        
+        try:
+            # Open image
+            image = Image.open(img_path).convert("RGB")
+            
+            # Load and parse annotation
+            with open(ann_path, 'r') as f:
+                ann_data = json.load(f)
+            
+            # Get boxes and labels from JSON
+            boxes = ann_data[ObjectDetectionKeys.BOXES]  # Expected: [[x1, y1, x2, y2], ...]
+            labels = ann_data[ObjectDetectionKeys.LABELS] # Expected: [1, 2, 1, ...]
+            
+            # Convert to tensors
+            target: Dict[str, Any] = {}
+            target[ObjectDetectionKeys.BOXES] = torch.as_tensor(boxes, dtype=torch.float32)
+            target[ObjectDetectionKeys.LABELS] = torch.as_tensor(labels, dtype=torch.int64)
+            
+        except Exception as e:
+            _LOGGER.error(f"Error loading sample #{idx}: {img_path.name} / {ann_path.name}. Error: {e}")
+            # Return empty/dummy data
+            return torch.empty(3, 224, 224), {ObjectDetectionKeys.BOXES: torch.empty((0, 4)), ObjectDetectionKeys.LABELS: torch.empty(0, dtype=torch.long)}
+
+        if self.transform:
+            image, target = self.transform(image, target)
+            
+        return image, target
+
+# Internal Paired Transform Helpers for Object Detection
+class _OD_PairedCompose:
+    """A 'Compose' for paired image/target_dict transforms."""
+    def __init__(self, transforms: List[Callable]):
+        self.transforms = transforms
+
+    def __call__(self, image: Any, target: Any) -> Tuple[Any, Any]:
+        for t in self.transforms:
+            image, target = t(image, target)
+        return image, target
+
+class _OD_PairedToTensor:
+    """Converts a PIL Image to Tensor, passes targets dict through."""
+    def __call__(self, image: Image.Image, target: Dict[str, Any]) -> Tuple[torch.Tensor, Dict[str, Any]]:
+        return TF.to_tensor(image), target
+
+class _OD_PairedNormalize:
+    """Normalizes the image tensor and leaves the target dict untouched."""
+    def __init__(self, mean: List[float], std: List[float]):
+        self.normalize = transforms.Normalize(mean, std)
+    
+    def __call__(self, image: torch.Tensor, target: Dict[str, Any]) -> Tuple[torch.Tensor, Dict[str, Any]]:
+        image_normalized = self.normalize(image)
+        return image_normalized, target
+
+class _OD_PairedRandomHorizontalFlip:
+    """Applies the same random horizontal flip to both image and targets['boxes']."""
+    def __init__(self, p: float = 0.5):
+        self.p = p
+    
+    def __call__(self, image: Image.Image, target: Dict[str, Any]) -> Tuple[Image.Image, Dict[str, Any]]:
+        if random.random() < self.p:
+            w, h = image.size
+            # Use new variable names to avoid linter confusion
+            flipped_image = TF.hflip(image) # type: ignore
+            
+            # Flip boxes
+            boxes = target[ObjectDetectionKeys.BOXES].clone() # [N, 4]
+            
+            # xmin' = w - xmax
+            # xmax' = w - xmin
+            boxes[:, 0] = w - target[ObjectDetectionKeys.BOXES][:, 2] # xmin'
+            boxes[:, 2] = w - target[ObjectDetectionKeys.BOXES][:, 0] # xmax'
+            target[ObjectDetectionKeys.BOXES] = boxes
+            
+            return flipped_image, target # type: ignore
+            
+        return image, target
+
+
+class ObjectDetectionDatasetMaker(_BaseMaker):
+    """
+    Creates processed PyTorch datasets for object detection from image
+    and JSON annotation folders.
+
+    This maker finds all matching image-annotation pairs from two directories,
+    splits them, and applies identical transformations (including augmentations)
+    to both the image and its corresponding target dictionary.
+    
+    The `DragonFastRCNN` model expects a list of images and a list of targets,
+    so this class provides a `collate_fn` to be used with a DataLoader.
+    
+    Workflow:
+    1. `maker = ObjectDetectionDatasetMaker.from_folders(img_dir, ann_dir)`
+    2. `maker.set_class_map({'background': 0, 'person': 1, 'car': 2})`
+    3. `maker.split_data(val_size=0.2)`
+    4. `maker.configure_transforms()`
+    5. `train_ds, val_ds = maker.get_datasets()`
+    6. `collate_fn = maker.collate_fn`
+    7. `train_loader = DataLoader(train_ds, ..., collate_fn=collate_fn)`
+    """
+    IMG_EXTENSIONS = ('.jpg', '.jpeg', '.png', '.bmp', '.tif', '.tiff')
+    
+    def __init__(self):
+        """
+        Typically not called directly. Use the class method `from_folders()` to create an instance.
+        """
+        super().__init__()
+        self.image_paths: List[Path] = []
+        self.annotation_paths: List[Path] = []
+        self.class_map: Dict[str, int] = {}
+        
+        self._is_split = False
+        self._are_transforms_configured = False
+        self.train_transform: Optional[Callable] = None
+        self.val_transform: Optional[Callable] = None
+        self._val_recipe_components: Optional[Dict[str, Any]] = None
+
+    @classmethod
+    def from_folders(cls, image_dir: Union[str, Path], annotation_dir: Union[str, Path]) -> 'ObjectDetectionDatasetMaker':
+        """
+        Creates a maker instance by loading all matching image-annotation pairs
+        from two corresponding directories.
+        
+        This method assumes that for an image `images/img_001.png`, there
+        is a corresponding annotation `annotations/img_001.json`.
+        
+        The JSON file must contain "boxes" and "labels" keys:
+        `{"boxes": [[x1,y1,x2,y2], ...], "labels": [1, 2, ...]}`
+
+        Args:
+            image_dir (str | Path): Path to the directory containing input images.
+            annotation_dir (str | Path): Path to the directory containing .json
+                                         annotation files.
+
+        Returns:
+            ObjectDetectionDatasetMaker: A new instance with all pairs loaded.
+        """
+        maker = cls()
+        img_path_obj = make_fullpath(image_dir, enforce="directory")
+        ann_path_obj = make_fullpath(annotation_dir, enforce="directory")
+
+        # Find all images
+        image_files = sorted([
+            p for p in img_path_obj.glob('*.*') 
+            if p.suffix.lower() in cls.IMG_EXTENSIONS
+        ])
+        
+        if not image_files:
+            _LOGGER.error(f"No images with extensions {cls.IMG_EXTENSIONS} found in {image_dir}")
+            raise FileNotFoundError()
+
+        _LOGGER.info(f"Found {len(image_files)} images. Searching for matching .json annotations in {annotation_dir}...")
+        
+        good_img_paths = []
+        good_ann_paths = []
+
+        for img_file in image_files:
+            # Find annotation with same stem + .json
+            ann_file = ann_path_obj / (img_file.stem + ".json")
+            
+            if ann_file.exists():
+                good_img_paths.append(img_file)
+                good_ann_paths.append(ann_file)
+            else:
+                _LOGGER.warning(f"No corresponding .json annotation found for image: {img_file.name}")
+        
+        if not good_img_paths:
+            _LOGGER.error("No matching image-annotation pairs were found.")
+            raise FileNotFoundError()
+            
+        _LOGGER.info(f"Successfully found {len(good_img_paths)} image-annotation pairs.")
+        maker.image_paths = good_img_paths
+        maker.annotation_paths = good_ann_paths
+        
+        return maker
+
+    @staticmethod
+    def inspect_folder(path: Union[str, Path]):
+        """
+        Logs a report of the types, sizes, and channels of image files
+        found in the directory.
+        """
+        VisionDatasetMaker.inspect_folder(path)
+
+    def set_class_map(self, class_map: Dict[str, int]) -> 'ObjectDetectionDatasetMaker':
+        """
+        Sets a map of class_name -> pixel_value. This is used by the
+        MLTrainer for clear evaluation reports.
+        
+        **Important:** For object detection models, 'background' MUST
+        be included as class 0.
+        Example: `{'background': 0, 'person': 1, 'car': 2}`
+
+        Args:
+            class_map (Dict[str, int]): A dictionary mapping the string name
+                to its integer label.
+        """
+        if 'background' not in class_map or class_map['background'] != 0:
+            _LOGGER.warning("Object detection class map should include 'background' mapped to 0.")
+        
+        self.class_map = class_map
+        _LOGGER.info(f"Class map set: {class_map}")
+        return self
+    
+    @property
+    def classes(self) -> List[str]:
+        """Returns the list of class names, if set."""
+        if self.class_map:
+            return list(self.class_map.keys())
+        return []
+
+    def split_data(self, val_size: float = 0.2, test_size: float = 0.0, 
+                   random_state: Optional[int] = 42) -> 'ObjectDetectionDatasetMaker':
+        """
+        Splits the loaded image-annotation pairs into train, validation, and test sets.
+
+        Args:
+            val_size (float): Proportion of the dataset to reserve for validation.
+            test_size (float): Proportion of the dataset to reserve for testing.
+            random_state (int | None): Seed for reproducible splits.
+
+        Returns:
+            ObjectDetectionDatasetMaker: The same instance, now with datasets created.
+        """
+        if self._is_split:
+            _LOGGER.warning("Data has already been split.")
+            return self
+
+        if val_size + test_size >= 1.0:
+            _LOGGER.error("The sum of val_size and test_size must be less than 1.")
+            raise ValueError()
+        
+        if not self.image_paths:
+            _LOGGER.error("There is no data to split. Use .from_folders() first.")
+            raise RuntimeError()
+        
+        indices = list(range(len(self.image_paths)))
+
+        # Split indices
+        train_indices, val_test_indices = train_test_split(
+            indices, test_size=(val_size + test_size), random_state=random_state
+        )
+        
+        # Helper to get paths from indices
+        def get_paths(idx_list):
+            return [self.image_paths[i] for i in idx_list], [self.annotation_paths[i] for i in idx_list]
+
+        train_imgs, train_anns = get_paths(train_indices)
+        
+        if test_size > 0:
+            val_indices, test_indices = train_test_split(
+                val_test_indices, test_size=(test_size / (val_size + test_size)), 
+                random_state=random_state
+            )
+            val_imgs, val_anns = get_paths(val_indices)
+            test_imgs, test_anns = get_paths(test_indices)
+            
+            self._test_dataset = _ObjectDetectionDataset(test_imgs, test_anns, transform=None)
+            self._test_dataset.classes = self.classes # type: ignore
+            _LOGGER.info(f"Test set created with {len(self._test_dataset)} images.")
+        else:
+            val_imgs, val_anns = get_paths(val_test_indices)
+        
+        self._train_dataset = _ObjectDetectionDataset(train_imgs, train_anns, transform=None)
+        self._val_dataset = _ObjectDetectionDataset(val_imgs, val_anns, transform=None)
+        
+        # Propagate class names to datasets for MLTrainer
+        self._train_dataset.classes = self.classes # type: ignore
+        self._val_dataset.classes = self.classes # type: ignore
+
+        self._is_split = True
+        _LOGGER.info(f"Data split into: \n- Training: {len(self._train_dataset)} images \n- Validation: {len(self._val_dataset)} images")
+        return self
+
+    def configure_transforms(self, 
+                             mean: List[float] = [0.485, 0.456, 0.406], 
+                             std: List[float] = [0.229, 0.224, 0.225]) -> 'ObjectDetectionDatasetMaker':
+        """
+        Configures and applies the image and target transformations.
+        
+        This method must be called AFTER data is split.
+        
+        For object detection models like Faster R-CNN, images are NOT
+        resized or cropped, as the model handles variable input sizes.
+        Transforms are limited to augmentation (flip), ToTensor, and Normalize.
+
+        Args:
+            mean (List[float]): The mean values for image normalization.
+            std (List[float]): The std dev values for image normalization.
+
+        Returns:
+            ObjectDetectionDatasetMaker: The same instance, with transforms applied.
+        """
+        if not self._is_split:
+            _LOGGER.error("Transforms must be configured AFTER splitting data. Call .split_data() first.")
+            raise RuntimeError()
+        
+        # --- Store components for validation recipe ---
+        self._val_recipe_components = {
+            VisionTransformRecipeKeys.MEAN: mean,
+            VisionTransformRecipeKeys.STD: std
+        }
+
+        # --- Validation/Test Pipeline (Deterministic) ---
+        self.val_transform = _OD_PairedCompose([
+            _OD_PairedToTensor(),
+            _OD_PairedNormalize(mean, std)
+        ])
+        
+        # --- Training Pipeline (Augmentation) ---
+        self.train_transform = _OD_PairedCompose([
+            _OD_PairedRandomHorizontalFlip(p=0.5),
+            _OD_PairedToTensor(),
+            _OD_PairedNormalize(mean, std)
+        ])
+
+        # --- Apply Transforms to the Datasets ---
+        self._train_dataset.transform = self.train_transform # type: ignore
+        self._val_dataset.transform = self.val_transform # type: ignore
+        if self._test_dataset:
+            self._test_dataset.transform = self.val_transform # type: ignore
+        
+        self._are_transforms_configured = True
+        _LOGGER.info("Paired object detection transforms configured and applied.")
+        return self
+
+    def get_datasets(self) -> Tuple[Dataset, ...]:
+        """
+        Returns the final train, validation, and optional test datasets.
+        
+        Raises:
+            RuntimeError: If called before data is split.
+            RuntimeError: If called before transforms are configured.
+        """
+        if not self._is_split:
+            _LOGGER.error("Data has not been split. Call .split_data() first.")
+            raise RuntimeError()
+        if not self._are_transforms_configured:
+            _LOGGER.error("Transforms have not been configured. Call .configure_transforms() first.")
+            raise RuntimeError()
+
+        if self._test_dataset:
+            return self._train_dataset, self._val_dataset, self._test_dataset
+        return self._train_dataset, self._val_dataset
+    
+    @property
+    def collate_fn(self) -> Callable:
+        """
+        Returns the collate function required by a DataLoader for this
+        dataset. This function ensures that images and targets are
+        batched as separate lists.
+        """
+        return _od_collate_fn
+    
+    def save_transform_recipe(self, filepath: Union[str, Path]) -> None:
+        """
+        Saves the validation transform pipeline as a JSON recipe file.
+        
+        For object detection, this recipe only includes ToTensor and
+        Normalize, as resizing is handled by the model.
+
+        Args:
+            filepath (str | Path): The path to save the .json recipe file.
+        """
+        if not self._are_transforms_configured:
+            _LOGGER.error("Transforms are not configured. Call .configure_transforms() first.")
+            raise RuntimeError()
+        
+        components = self._val_recipe_components
+        
+        if not components:
+            _LOGGER.error(f"Error getting the transformers recipe for validation set.")
+            raise ValueError()
+        
+        # validate path
+        file_path = make_fullpath(filepath, make=True, enforce="file")
+
+        # Add standard transforms
+        recipe: Dict[str, Any] = {
+            VisionTransformRecipeKeys.TASK: "object_detection",
+            VisionTransformRecipeKeys.PIPELINE: [
+                {VisionTransformRecipeKeys.NAME: "ToTensor", "kwargs": {}},
+                {VisionTransformRecipeKeys.NAME: "Normalize", "kwargs": {
+                    "mean": components["mean"],
+                    "std": components["std"]
+                }}
+            ]
+        }
+        
+        # Save the file
+        save_recipe(recipe, file_path)
+
+
+def info():
+    _script_info(__all__)