generate_occluded_imagenet.py

import os
import random
import numpy as np
import pandas as pd
from PIL import Image
from torchvision import datasets, transforms, io


def get_random_occluder(dataset_occluder):
    index = random.randint(0, len(dataset_occluder) - 1)
    texture_path, texture_class_index = dataset_occluder.imgs[index]
    texture_class = dataset_occluder.classes[texture_class_index]

    # load with the alpha channel
    texture = io.read_image(texture_path, mode=io.image.ImageReadMode.RGB_ALPHA)

    return texture, texture_class


def resize_occluder(occluder_pil, target_area):
    alpha = np.array(occluder_pil.getchannel('A'))
    non_transparent_area = np.count_nonzero(alpha > 0)

    area_scale_factor = target_area / non_transparent_area

    width_scale_factor = np.sqrt(area_scale_factor * (occluder_pil.width / occluder_pil.height))
    height_scale_factor = np.sqrt(area_scale_factor * (occluder_pil.height / occluder_pil.width))

    new_width = occluder_pil.width * width_scale_factor
    new_height = occluder_pil.height * height_scale_factor

    resized_occluder = occluder_pil.resize((int(new_width), int(new_height)), Image.LANCZOS)

    return resized_occluder


def randomly_rotate_occluder(occluder_pil):
    angle = random.uniform(-180, 180)
    return occluder_pil.rotate(angle, resample=Image.BICUBIC, expand=True)


def calculate_distance(point1, point2):
    x1, y1 = point1
    x2, y2 = point2
    return ((x2 - x1) ** 2 + (y2 - y1) ** 2) ** 0.5


def try_rotations(occluder_pil, image_pil, target_area, pos1=None):
    best_occluder = None
    best_area = 0
    best_pos = None
    min_distance = 130 # min distance if two occluders
    for _ in range(75):  # increase number of attempts to find better position
        rotated = randomly_rotate_occluder(occluder_pil)
        resized = resize_occluder(rotated, target_area)
        if resized.width > image_pil.width or resized.height > image_pil.height:
            pos = (image_pil.width // 2 - resized.width // 2, 
                   image_pil.height // 2 - resized.height // 2)
        else:
            max_x = max(0, image_pil.width - resized.width)
            max_y = max(0, image_pil.height - resized.height)
            pos = (random.randint(0, max_x), random.randint(0, max_y))

        if pos1 is not None and calculate_distance(pos1, pos) < min_distance:
            continue

        mask = Image.new('1', image_pil.size)
        mask.paste(resized.getchannel('A'), pos, resized.getchannel('A'))
        area = np.count_nonzero(np.array(mask))

        if area > best_area:
            best_area = area
            best_occluder = resized
            best_pos = pos
          
    return best_occluder, best_pos


def occlude_image(image, occluder_tensor, percentage_occlusion, occluded_dir, img_name):
    occluder_pil = transforms.ToPILImage(mode='RGBA')(occluder_tensor)
    image_pil = transforms.ToPILImage()(image)

    binary_mask = Image.new('1', image_pil.size)
    
    rn = random.random()

    if rn < 0.5:  # randomly use two occluders
        occluder_resizing_factor = 0.5
    else:
        occluder_resizing_factor = 1.0

    target_area = image_pil.width * image_pil.height * percentage_occlusion * occluder_resizing_factor

    if rn < 0.5:  # randomly use two occluders (can make this k occluders)
        occluder_pil1, pos1 = try_rotations(occluder_pil, image_pil, target_area / 2)
        image_pil.paste(occluder_pil1, pos1, occluder_pil1)
        occluder_alpha1 = occluder_pil1.getchannel('A')
        binary_mask.paste(occluder_alpha1, pos1, occluder_alpha1)
        
        occluder_pil2, pos2 = try_rotations(occluder_pil, image_pil, target_area / 2, pos1)
        if occluder_pil2 is not None and pos2 is not None:
            image_pil.paste(occluder_pil2, pos2, occluder_pil2)
            occluder_alpha2 = occluder_pil2.getchannel('A')
            binary_mask.paste(occluder_alpha2, pos2, occluder_alpha2)
        
        if pos2 is None:
            pos = [pos1]
        else:
            pos = [pos1, pos2]
    else:
        occluder_pil, pos = try_rotations(occluder_pil, image_pil, target_area)
        image_pil.paste(occluder_pil, pos, occluder_pil)
        occluder_alpha = occluder_pil.getchannel('A')
        binary_mask.paste(occluder_alpha, pos, occluder_alpha)

        pos = [pos]

    image_with_occluder_tensor = transforms.ToTensor()(image_pil)

    mask_array = np.array(binary_mask)
    mask_path = os.path.join(occluded_dir, f"{img_name}_mask.npy")
    np.save(mask_path, mask_array)

    return image_with_occluder_tensor, mask_path, pos


def rebuild_display_mask(image_path, mask_path):
    image_pil = Image.open(image_path)
    binary_mask = Image.new('1', image_pil.size)

    mask_array = np.load(mask_path)
    mask_indices = np.transpose(np.nonzero(mask_array))

    for i, j in mask_indices:
        binary_mask.putpixel((j, i), 1)

    binary_mask.show()


def build_dataset(data_path, transform):
    dataset = datasets.ImageFolder(data_path, transform=transform)
    nb_classes = len(dataset.classes)
    return dataset, nb_classes


def build_transform():
    t = []
    t.append(transforms.ToTensor())
    return transforms.Compose(t)


def main():
    data_dir = 'imagenet'
    texture_dir = 'occluders_segmented'
    occluded_data_dir = 'imagenet_occluded'

    transform = build_transform()
    dataset, nb_classes = build_dataset(data_dir, transform)
    dataset_occluder, _ = build_dataset(texture_dir, transform)
    percentage_occlusion = 0.3 # ~30%

    occlusion_info = pd.DataFrame(columns=["image_name", "class_name", "occluder_class",
                                           "percentage_occlusion", "mask", "pos"])
    
    count = 0

    for idx in range(len(dataset)):
        image, label = dataset[idx]
        category = dataset.classes[label]

        in_dir = os.path.join(data_dir, category)
        occluded_dir = os.path.join(occluded_data_dir, category)

        os.makedirs(occluded_dir, exist_ok=True)

        img_name = dataset.imgs[idx][0].split('/')[-1].split('.')[0]

        occluder_tensor, occluder_class = get_random_occluder(dataset_occluder)
        occluded_image, mask_path, pos = occlude_image(image, occluder_tensor, 
                                                       percentage_occlusion,
                                                       occluded_dir,
                                                       img_name)

        mask_array = np.load(mask_path)
        actual_percentage_occlusion = np.count_nonzero(mask_array) / (image.shape[1] * image.shape[2])

        occluded_image_path = os.path.join(occluded_dir, f"{img_name}_occluded.png")
        transforms.ToPILImage()(occluded_image).save(occluded_image_path)

        new_row = pd.DataFrame({
            "image_name": [f"{img_name}_occluded.png"],
            "class_name": [category],
            "occluder_class": [occluder_class],
            "percentage_occlusion": [actual_percentage_occlusion],
            "mask": [mask_path],
            "pos": [pos]
        })

        occlusion_info = pd.concat([occlusion_info, new_row], ignore_index=True)
        if count % 50 == 0: 
            print("Folder: {}/1000 ({})".format(count / 50, category))
        count+=1

    occlusion_info.to_csv(os.path.join(occluded_data_dir, "occlusion_info.csv"), index=False)

if __name__ == "__main__":
    main()