load_and_process_coco.py

import os
import tensorflow as tf
import tensorflow_datasets as tfds
import matplotlib.pyplot as plt
from datasets import Dataset, Features, Image, Sequence, Value

# --- Configuration ---
TARGET_SIZE = 32  # downsample the images to TARGET_SIZE x TARGET_SIZE
MIN_BOX_AREA_PIXELS = int(0.85*TARGET_SIZE)**2  # minimum area in px, that a depicted person must occupy to be included in the dataset.
CHANNELS = 3  # use 1 for grayscale
OUTLIER_PERCENTAGE = 0.001  # 2% of images will become extreme outliers


def crop_to_bbox_center(image, boxes, labels, target_size=TARGET_SIZE):
    """
    Crops an image based on the largest person bounding box.
    If no person is present, performs a default central crop.

    Args:
        image: 3D Tensor (H, W, C)
        boxes: 2D Tensor (N, 4) in normalized coordinates [ymin, xmin, ymax, xmax]
        labels: 1D Tensor (N,) containing class IDs (0 for person)
        target_size: Int, the final resized dimension
    """
    image_shape = tf.shape(image)
    h = tf.cast(image_shape[0], tf.float32)
    w = tf.cast(image_shape[1], tf.float32)

    # 1. Mask for 'person' class (label == 0)
    person_mask = tf.equal(labels, 0)
    has_person = tf.reduce_any(person_mask)

    def crop_with_person():
        # Filter boxes to only include people
        person_boxes = tf.boolean_mask(boxes, person_mask)

        # Calculate areas: (ymax - ymin) * (xmax - xmin)
        box_heights = person_boxes[:, 2] - person_boxes[:, 0]
        box_widths = person_boxes[:, 3] - person_boxes[:, 1]
        areas = box_heights * box_widths

        # Find the index of the largest person box
        largest_idx = tf.argmax(areas)
        best_box = person_boxes[largest_idx]

        # Convert normalized coordinates to absolute pixels
        ymin = best_box[0] * h
        xmin = best_box[1] * w
        ymax = best_box[2] * h
        xmax = best_box[3] * w

        # Determine square crop size based on the largest dimension of the box
        side = tf.maximum(ymax - ymin, xmax - xmin)
        # Ensure side isn't larger than the image canvas itself
        side = tf.minimum(side, tf.minimum(h, w))

        # Calculate center coordinates of the target box
        center_y = (ymin + ymax) / 2.0
        center_x = (xmin + xmax) / 2.0

        # Calculate top-left corner of the crop square
        crop_ymin = center_y - (side / 2.0)
        crop_xmin = center_x - (side / 2.0)

        # Clamp to ensure the crop window stays completely inside the image boundaries
        crop_ymin = tf.clip_by_value(crop_ymin, 0.0, h - side)
        crop_xmin = tf.clip_by_value(crop_xmin, 0.0, w - side)

        return crop_ymin, crop_xmin, side

    def crop_anywhere():
        # Fallback if no person: default to a center crop spanning the shortest side
        side = tf.minimum(h, w)
        crop_ymin = (h - side) / 2.0
        crop_xmin = (w - side) / 2.0
        return crop_ymin, crop_xmin, side

    # Dynamically choose the cropping window coordinates based on condition
    crop_ymin, crop_xmin, side = tf.cond(has_person, crop_with_person, crop_anywhere)

    # Perform the physical crop
    cropped_image = tf.image.crop_to_bounding_box(
        image,
        tf.cast(crop_ymin, tf.int32),
        tf.cast(crop_xmin, tf.int32),
        tf.cast(side, tf.int32),
        tf.cast(side, tf.int32)
    )

    # Resize to final target dimensions
    final_image = tf.image.resize(cropped_image, [target_size, target_size])

    return final_image, boxes


def prepare_data(example):
    """ Apply cropping and preserve labels. """
    image = example['image']
    boxes = example['objects']['bbox']
    labels = example['objects']['label']
    image, boxes = crop_to_bbox_center(image, boxes, labels)

    # Set color space
    if CHANNELS == 1:
        image = tf.image.rgb_to_grayscale(image)
    elif CHANNELS == 3:
        pass
    else:
        raise ValueError(f"Unsupported CHANNELS value: {CHANNELS}. Expected 1 or 3.")

    image = tf.cast(image, tf.uint8)
    labels = tf.cast(labels, tf.uint8)
    return {'image': image, 'bbox': boxes, 'label': labels}


def filter_small_persons(example, target_size=TARGET_SIZE, min_box_area=MIN_BOX_AREA_PIXELS):
    """
    Keep image if AT LEAST ONE person is larger than the threshold post-crop.
    If the image contains no people at all, keep it as is.
    """
    labels = example['objects']['label']
    boxes = example['objects']['bbox']  # [ymin, xmin, ymax, xmax] (normalized)

    is_person = tf.equal(labels, 0)
    person_boxes = tf.boolean_mask(boxes, is_person)
    has_persons = tf.greater(tf.shape(person_boxes)[0], 0)

    def process_person_images():
        image_shape = tf.shape(example['image'])
        h = tf.cast(image_shape[0], tf.float32)
        w = tf.cast(image_shape[1], tf.float32)

        # 1. Calculate individual absolute box areas to find the single largest one
        box_heights_normalized = person_boxes[:, 2] - person_boxes[:, 0]
        box_widths_normalized = person_boxes[:, 3] - person_boxes[:, 1]

        # Absolute dimensions for area calculation
        box_heights_orig = box_heights_normalized * h
        box_widths_orig = box_widths_normalized * w
        areas_orig = box_heights_orig * box_widths_orig

        # 2. Replicate the NEW cropping logic: Find the SINGLE largest person box
        largest_idx = tf.argmax(areas_orig)
        best_box = person_boxes[largest_idx]

        # Convert the single largest box to absolute pixels
        ymin = best_box[0] * h
        xmin = best_box[1] * w
        ymax = best_box[2] * h
        xmax = best_box[3] * w

        # The side length depends ONLY on this largest person box
        side = tf.maximum(ymax - ymin, xmax - xmin)
        side = tf.minimum(side, tf.minimum(h, w))

        # 3. Project ALL person boxes into the final target resolution based on this side length
        scale_factor = target_size / side
        final_box_heights = box_heights_orig * scale_factor
        final_box_widths = box_widths_orig * scale_factor

        # Calculate final resolution pixel areas
        final_box_areas = final_box_heights * final_box_widths

        # Check if any individual person meets or exceeds the threshold
        is_big_enough = final_box_areas >= min_box_area

        # Keep the image if AT LEAST ONE person is big enough post-crop
        return tf.reduce_any(is_big_enough)

    # If there are no people, return True to keep the image untouched.
    return tf.cond(has_persons, process_person_images, lambda: tf.constant(True))


def visualize_dataset(dataset, num_images=20):
    """ Display a portion of random images and person images. """
    # Top row: Grab the first 10 processed images sequentially
    top_row_ds = dataset.take(num_images)

    # Bottom row: Since labels are preserved, we filter for elements containing a person (0)
    bottom_row_ds = dataset.filter(lambda x: tf.reduce_any(tf.equal(x['label'], 0))).take(num_images)

    fig, axes = plt.subplots(2, num_images, figsize=(18, 5))

    # Row 1: Random Samples
    for i, example in enumerate(top_row_ds):
        img = example['image']

        axes[0, i].imshow(img)
        axes[0, i].set_title(f"Random {i + 1}", fontsize=9)
        axes[0, i].axis('off')

    # Row 2: Person Samples
    for i, example in enumerate(bottom_row_ds):
        img = example['image']
        axes[1, i].imshow(img)
        axes[1, i].set_title(f"Person {i + 1}", fontsize=9)
        axes[1, i].axis('off')

    plt.tight_layout()
    plt.show()


def save_dataset_to_hf_format(tf_dataset, target_size, export_dir=None):
    """
    Converts a processed tf.data.Dataset into a universally accessible
    Hugging Face Parquet dataset on disk.
    """
    print("Converting TensorFlow dataset to standard arrays...")

    images_list = []
    bboxes_list = []
    labels_list = []

    # Stream the elements out of the TF pipeline into memory
    for example in tf_dataset:
        images_list.append(example['image'].numpy())
        bboxes_list.append(example['bbox'].numpy().tolist())
        labels_list.append(example['label'].numpy().tolist())

    # Define standard HF schema so the Hub can preview the images natively
    features = Features({
        'image': Image(),
        'bbox': Sequence(Sequence(Value('float32'))),
        'label': Sequence(Value('int64'))
    })

    # Build the Hugging Face dataset structure
    hf_dataset = Dataset.from_dict({
        'image': images_list,
        'bbox': bboxes_list,
        'label': labels_list
    }, features=features)

    # Determine the save path
    if export_dir is None:
        export_dir = os.path.expanduser(f'~/Documents/local/data/tiny_presence_{target_size}_hf')
    else:
        export_dir = os.path.expanduser(export_dir)

    os.makedirs(export_dir, exist_ok=True)

    # Save to disk as highly optimized, cross-framework Parquet files
    print(f"Saving universally compatible dataset to: {export_dir}")
    hf_dataset.save_to_disk(export_dir)
    print("Dataset successfully saved!")
    return hf_dataset


def visualize_outliers(outlier_samples, max_display=10):
    """ Renders the detected outlier samples in a single row layout. """
    if not outlier_samples:
        print("No outlier samples collected to visualize.")
        return

    n_display = min(len(outlier_samples), max_display)
    fig, axes = plt.subplots(1, n_display, figsize=(18, 3))

    # Ensure axes is iterable even if there's only 1 image
    if n_display == 1:
        axes = [axes]

    for i in range(n_display):
        img, img_mean = outlier_samples[i]

        if CHANNELS == 1:
            axes[i].imshow(img.squeeze(), cmap='gray')
        else:
            axes[i].imshow(img)

        axes[i].set_title(f"Mean: {img_mean:.1f}", fontsize=9)
        axes[i].axis('off')

    plt.suptitle(f"Sample Visualized Outliers (Target Size: {TARGET_SIZE}x{TARGET_SIZE})", fontsize=12, weight='bold')
    plt.tight_layout()
    plt.show()


def introduce_outliers(example, outlier_prob=OUTLIER_PERCENTAGE):
    """ Randomly alters images to be either extremely bright or extremely dark. """
    image = example['image']
    rand_val = tf.random.uniform([], minval=0.0, maxval=1.0)
    should_be_outlier = rand_val < outlier_prob

    def make_outlier():
        outlier_type = tf.random.uniform([], minval=0, maxval=2, dtype=tf.int32)
        image_float = tf.cast(image, tf.float32)

        outlier_skew = tf.random.uniform([], minval=200.0, maxval=210.0)

        modified_image = tf.cond(
            tf.equal(outlier_type, 0),
            lambda: image_float + outlier_skew,  # Ultra-bright
            lambda: image_float - outlier_skew   # Ultra-dark
        )
        modified_image = tf.clip_by_value(modified_image, 0.0, 255.0)
        return tf.cast(modified_image, tf.uint8)

    final_image = tf.cond(should_be_outlier, make_outlier, lambda: image)

    return {
        'image': final_image,
        'bbox': example['bbox'],
        'label': example['label'],
        'is_outlier': should_be_outlier
    }


# --- DATA PROCESSING PIPELINE ---
# Load data
ds_builder = tfds.builder('coco/2017', data_dir='~/Documents/local/data')
ds_train = ds_builder.as_dataset(split='train')
ds_val = ds_builder.as_dataset(split='validation')
ds_test = ds_builder.as_dataset(split='test')

# Join datasets
full_ds = ds_train.concatenate(ds_val).concatenate(ds_test)
n_images_raw = sum(1 for _ in full_ds)

# Apply filters + cropping
processed_ds = full_ds.filter(lambda x: tf.shape(x['objects']['label'])[0] > 0)
processed_ds = processed_ds.filter(filter_small_persons)
processed_ds = processed_ds.map(prepare_data, num_parallel_calls=tf.data.AUTOTUNE)

# Create outliers
processed_ds = processed_ds.map(introduce_outliers, num_parallel_calls=tf.data.AUTOTUNE)

# Print statistics & collect outlier visualization arrays
n_images_processed = 0
n_outliers = 0
outlier_pixel_sum = 0.0
collected_outliers = []  # Holds tuples of (image_numpy, mean_value)
MAX_SAMPLES_TO_COLLECT = 10

print("Analyzing dataset and collecting metrics...")
for example in processed_ds:
    n_images_processed += 1

    if example['is_outlier'].numpy():
        n_outliers += 1
        img_np = example['image'].numpy()
        current_mean = img_np.mean()
        outlier_pixel_sum += current_mean

        # Save up to MAX_SAMPLES_TO_COLLECT items to plot later
        if len(collected_outliers) < MAX_SAMPLES_TO_COLLECT:
            collected_outliers.append((img_np, current_mean))

avg_outlier_pixel_val = (outlier_pixel_sum / n_outliers) if n_outliers > 0 else 0.0
outlier_ratio = (n_outliers / n_images_processed * 100) if n_images_processed > 0 else 0.0

print(f"\n--- Dataset Statistics ---")
print(f"Total images in raw dataset: {n_images_raw}")
print(f"Images discarded (filtered): {n_images_raw - n_images_processed}")
print(f"Images remaining:            {n_images_processed}")
print(f"--------------------------")
print(f"--- Outlier Statistics ---")
print(f"Detected Outliers:           {n_outliers} ({outlier_ratio:.2f}% of remaining)")
print(f"Average Outlier Pixel Mean:  {avg_outlier_pixel_val:.2f} (Expected near 0 or 255)")
print(f"--------------------------")

# Visualize original/person subsets
visualize_dataset(processed_ds)

# Visualize the captured outlier images
visualize_outliers(collected_outliers, max_display=MAX_SAMPLES_TO_COLLECT)

# Size estimation
total_bytes = TARGET_SIZE * TARGET_SIZE * CHANNELS * tf.uint8.size * n_images_processed
print(f"Total dataset cache size: {total_bytes / (1024 * 1024):.2f} MB")

# Save
processed_ds = processed_ds.apply(tf.data.experimental.assert_cardinality(n_images_processed))
corr_label = "_with_corruptions" if OUTLIER_PERCENTAGE else ""
save_path = os.path.expanduser(f'~/Documents/local/data/tiny_presence_{TARGET_SIZE}_CH_{CHANNELS}_easy{corr_label}')
os.makedirs(save_path, exist_ok=True)
tf.data.Dataset.save(processed_ds, save_path, compression='GZIP')
print(f"Dataset successfully saved to: {save_path}")