scripts/extract_crop_lips.py

import sys

sys.path.append('..')

import os.path
import options

import cv2
import dlib
import numpy as np
import options as opt
import matplotlib.pyplot as plt

from tqdm.auto import tqdm
from multiprocessing import Pool

predictor_path = '../pretrain/shape_predictor_68_face_landmarks.dat'
predictor = dlib.shape_predictor(predictor_path)
detector = dlib.get_frontal_face_detector()

RUN_PARALLEL = True
FORCE_RATIO = True
BORDER = 10

base = os.path.abspath('..')
image_dir = os.path.join(base, options.images_dir)
anno_dir = os.path.join(base, options.alignments_dir)
crop_dir = os.path.join(base, options.crop_images_dir)


def get_mouth_marks(shape):
    marks = np.zeros((2, 20))
    co = 0

    # Specific for the mouth.
    for ii in range(48, 68):
        """
        This for loop is going over all mouth-related features.
        X and Y coordinates are extracted and stored separately.
        """
        X = shape.part(ii)
        A = (X.x, X.y)
        marks[0, co] = X.x
        marks[1, co] = X.y
        co += 1

    # Get the extreme points(top-left & bottom-right)
    X_left, Y_left, X_right, Y_right = [
        int(np.amin(marks, axis=1)[0]),
        int(np.amin(marks, axis=1)[1]),
        int(np.amax(marks, axis=1)[0]),
        int(np.amax(marks, axis=1)[1])
    ]

    return X_left, Y_left, X_right, Y_right


translate_pairs = []

for speaker_no in range(1, 35):
    speaker_name = f's{speaker_no}'
    speaker_image_dir = os.path.join(image_dir, speaker_name)
    speaker_crop_dir = os.path.join(crop_dir, speaker_name)
    speaker_anno_dir = os.path.join(anno_dir, speaker_name)

    if not os.path.exists(speaker_image_dir):
        continue
    if not os.path.exists(speaker_crop_dir):
        os.mkdir(speaker_crop_dir)

    sentence_dirs = os.listdir(speaker_image_dir)

    for sentence in sentence_dirs:
        anno_filepath = os.path.join(speaker_anno_dir, f'{sentence}.align')
        if not os.path.exists(anno_filepath):
            continue

        translate_pairs.append((speaker_no, sentence))


print('PAIRS', len(translate_pairs))


def extract_mouth_image(speaker_no, sentence):
    speaker_name = f's{speaker_no}'
    speaker_image_dir = os.path.join(image_dir, speaker_name)
    speaker_crop_dir = os.path.join(crop_dir, speaker_name)

    img_sentence_dir = os.path.join(speaker_image_dir, sentence)
    crop_sentence_dir = os.path.join(speaker_crop_dir, sentence)
    filenames = os.listdir(img_sentence_dir)

    if not os.path.exists(crop_sentence_dir):
        os.mkdir(crop_sentence_dir)

    for filename in filenames:
        img_filepath = os.path.join(img_sentence_dir, filename)
        if not img_filepath.endswith('.jpg'):
            continue

        crop_filepath = os.path.join(crop_sentence_dir, filename)
        image = cv2.imread(img_filepath)

        detection_bbox = detector(image, 1)[0]
        # (360 x 288 x 3)
        width, height, depth = image.shape
        shape = predictor(image, detection_bbox)
        X_left, Y_left, X_right, Y_right = get_mouth_marks(shape)

        # Find the center of the mouth.
        X_center = (X_left + X_right) / 2.0
        Y_center = (Y_left + Y_right) / 2.0

        # Make a boarder for cropping.
        X_left_new = X_left - BORDER
        Y_left_new = Y_left - BORDER
        X_right_new = X_right + BORDER
        Y_right_new = Y_right + BORDER

        # Width and height for cropping
        # (before and after considering the border)
        width_new = X_right_new - X_left_new
        height_new = Y_right_new - Y_left_new
        width_current = X_right - X_left
        height_current = Y_right - Y_left

        height_crop_max = height_new
        width_crop_max = width_new

        if width_crop_max % 2 == 1:
            width_crop_max += 1
        if height_crop_max % 2 == 1:
            height_crop_max += 1

        if FORCE_RATIO:
            if width_crop_max < height_crop_max * 2:
                width_crop_max = height_crop_max * 2
            else:
                height_crop_max = width_crop_max // 2

        # Find the cropping points(top-left and bottom-right).
        X_left_crop = int(X_center - width_crop_max / 2.0)
        X_right_crop = int(X_center + width_crop_max / 2.0)
        Y_left_crop = int(Y_center - height_crop_max / 2.0)
        Y_right_crop = int(Y_center + height_crop_max / 2.0)
        X_left_crop = max(X_left_crop, 0)
        Y_left_crop = max(Y_left_crop, 0)

        mouth = image[
            Y_left_crop:Y_right_crop, X_left_crop:X_right_crop, :
        ]

        if FORCE_RATIO:
            height, width, _ = mouth.shape

            if width != height * 2:
                mouth = cv2.resize(
                    mouth, dsize=(height * 2, height),
                    interpolation=cv2.INTER_CUBIC
                )

        # print('SHAPE', mouth.shape)
        # print('IMG_PATH', img_filepath)
        # plt.imshow(mouth); plt.show()
        # print('CF', crop_filepath)
        cv2.imwrite(crop_filepath, mouth)

    return speaker_no, sentence


if RUN_PARALLEL:
    def kwargify(**kwargs): return kwargs


    pbar = tqdm(translate_pairs)
    pool = Pool(processes=12)
    jobs = []


    def callback(resp):
        pbar.desc = str(resp)
        pbar.update(1)


    for translate_pair in translate_pairs:
        speaker_no, sentence = translate_pair

        job_kwargs = kwargify(
            speaker_no=speaker_no, sentence=sentence
        )
        job = pool.apply_async(
            extract_mouth_image, kwds=job_kwargs,
            callback=callback
        )

        jobs.append(job)

    # Wait for all tasks to complete
    for job in jobs:
        job.wait()

    pool.close()
    pool.join()
else:
    for translate_pair in tqdm(translate_pairs):
        extract_mouth_image(*translate_pair)