Cards can be skewed now

Edmond Yoo

2018-09-07 391bff5362190e220b457f1be0cfb66ba706e71d

Cards can be skewed now

 transform_data.py

@@ -41,11 +41,12 @@
    """
    A template for generating a training image.
    """
    def __init__(self, img_bg, width, height, cards=None):
    def __init__(self, img_bg, width, height, skew=None, cards=None):
        """
        :param img_bg: background (textile) image
        :param width: width of the training image
        :param height: height of the training image
        :param skew: 4 coordinates that indicates the corners (in normalized form) for perspective transform
        :param cards: list of Card objects
        """
        self.img_bg = img_bg
@@ -56,6 +57,15 @@
            self.cards = []
        else:
            self.cards = cards

        # Compute transform matrix for perspective transform
        if skew is not None:
            orig_corner = np.array([[0, 0], [0, height], [width, height], [width, 0]], dtype=np.float32)
            new_corner = np.array([[width * s[0], height * s[1]] for s in skew], dtype=np.float32)
            self.M = cv2.getPerspectiveTransform(orig_corner, new_corner)
            pass
        else:
            self.M = None
        pass

    def add_card(self, card, x=None, y=None, theta=0.0, scale=1.0):
@@ -85,7 +95,8 @@
        :return: none
        """
        self.check_visibility(visibility=visibility)
        img_result = cv2.resize(self.img_bg, (self.width, self.height))
        #img_result = cv2.resize(self.img_bg, (self.width, self.height))
        img_result = np.zeros((self.height, self.width, 3), dtype=np.uint8)

        for card in self.cards:
            if card.x == 0.0 and card.y == 0.0 and card.theta == 0.0 and card.scale == 1.0:
@@ -125,9 +136,10 @@
                for ext_obj in card.objects:
                    if ext_obj.visible:
                        for pt in ext_obj.key_pts:
                            cv2.circle(img_result, card.coordinate_in_generator(pt[0], pt[1]), 2, (0, 0, 255), 2)
                            cv2.circle(img_result, card.coordinate_in_generator(pt[0], pt[1]), 2, (1, 1, 255), 10)
                        bounding_box = card.bb_in_generator(ext_obj.key_pts)
                        cv2.rectangle(img_result, bounding_box[0], bounding_box[2], (0, 255, 0), 2)
                        cv2.rectangle(img_result, bounding_box[0], bounding_box[2], (1, 255, 1), 5)

        '''
        try:
            text = pytesseract.image_to_string(img_result, output_type=pytesseract.Output.DICT)
@@ -137,6 +149,21 @@
        '''
        img_result = cv2.GaussianBlur(img_result, (5, 5), 0)

        if self.M is not None:
            img_result = cv2.warpPerspective(img_result, self.M, (self.width, self.height))
            if debug:
                for card in self.cards:
                    for ext_obj in card.objects:
                        if ext_obj.visible:
                            new_pts = np.array([[list(card.coordinate_in_generator(pt[0], pt[1]))]
                                                for pt in ext_obj.key_pts], dtype=np.float32)
                            new_pts = cv2.perspectiveTransform(new_pts, self.M)
                            for pt in new_pts:
                                cv2.circle(img_result, (pt[0][0], pt[0][1]), 2, (255, 1, 1), 10)

        img_bg = cv2.resize(self.img_bg, (self.width, self.height))
        img_result = np.where(img_result, img_result, img_bg)

        if aug is not None:
            img_result = aug.augment_image(img_result)

@@ -461,6 +488,7 @@
    ia.seed(random.randrange(10000))

    bg_images = generate_data.load_dtd(dtd_dir='%s/dtd/images' % data_dir, dump_it=False)
    #bg_images = [cv2.imread('data/frilly_0007.jpg')]
    background = generate_data.Backgrounds(images=bg_images)

    card_pool = pd.DataFrame()
@@ -472,7 +500,11 @@
    num_iter = 1

    for i in range(num_gen):
        generator = ImageGenerator(background.get_random(), 1440, 960)
        # Arbitrarily select top left and right corners for perspective transformation
        # Since the training image are generated with random rotation, don't need to skew all four sides
        skew = [[random.uniform(0, 0.25), 0], [0, 1], [1, 1],
                [random.uniform(0.75, 1), 0]]
        generator = ImageGenerator(background.get_random(), 1440, 960, skew=skew)
        out_name = ''
        for _, card_info in card_pool.sample(random.randint(2, 5)).iterrows():
            img_name = '%s/card_img/png/%s/%s_%s.png' % (data_dir, card_info['set'], card_info['collector_number'],
@@ -492,8 +524,8 @@
                iaa.Multiply((0.8, 1.2)),  # darken / brighten the whole image
                iaa.SimplexNoiseAlpha(first=iaa.Add(random.randrange(64)), per_channel=0.1, size_px_max=[3, 6],
                                      upscale_method="cubic"),  # Lighting
                iaa.AdditiveGaussianNoise(scale=random.uniform(0.005, 0.05) * 255, per_channel=0.1),  # Noises
                iaa.Dropout(p=[0.005, 0.05], per_channel=0.1)
                iaa.AdditiveGaussianNoise(scale=random.uniform(0, 0.05) * 255, per_channel=0.1),  # Noises
                iaa.Dropout(p=[0, 0.05], per_channel=0.1)
            ])
            if i % 3 == 0:
                generator.generate_non_obstructive()
@@ -507,6 +539,8 @@
                generator.generate_vertical_span(theta=random.uniform(-math.pi, math.pi))
                generator.export_training_data(visibility=0.0, out_name='%s/train/vertical_span/%s_%d'
                                                                        % (data_dir, out_name, j), aug=seq)
            #generator.generate_horizontal_span(theta=random.uniform(-math.pi, math.pi))
            #generator.render(display=True, aug=seq, debug=True)
            print('Generated %s%d' % (out_name, j))
            generator.img_bg = background.get_random()
    pass