Moving files from MTGCardDetector repo

Edmond Yoo

2018-09-16 0dab894a5be9f7d10d85e89dea91d02c71bae84d

 transform_data.py

@@ -1,3 +1,4 @@
import os
import random
import math
import cv2
@@ -8,8 +9,13 @@
import generate_data
from shapely import geometry
import pytesseract
import imgaug as ia
from imgaug import augmenters as iaa
from imgaug import parameters as iap

card_mask = cv2.imread('data/mask.png')
data_dir = os.path.abspath('/media/win10/data')
darknet_dir = os.path.abspath('darknet')


def key_pts_to_yolo(key_pts, w_img, h_img):
@@ -20,10 +26,10 @@
    :param h_img: height of the entire image
    :return: <x> <y> <width> <height>
    """
    x1 = min([pt[0] for pt in key_pts])
    x2 = max([pt[0] for pt in key_pts])
    y1 = min([pt[1] for pt in key_pts])
    y2 = max([pt[1] for pt in key_pts])
    x1 = max(0, min([pt[0] for pt in key_pts]))
    x2 = min(w_img, max([pt[0] for pt in key_pts]))
    y1 = max(0, min([pt[1] for pt in key_pts]))
    y2 = min(h_img, max([pt[1] for pt in key_pts]))
    x = (x2 + x1) / 2 / w_img
    y = (y2 + y1) / 2 / h_img
    width = (x2 - x1) / w_img
@@ -35,14 +41,16 @@
    """
    A template for generating a training image.
    """
    def __init__(self, img_bg, width, height, cards=None):
    def __init__(self, img_bg, class_ids, 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
        self.class_ids = class_ids
        self.img_result = None
        self.width = width
        self.height = height
@@ -50,6 +58,15 @@
            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):
@@ -73,13 +90,14 @@
        card.scale = scale
        pass

    def render(self, visibility=0.5, display=False, debug=False):
    def render(self, visibility=0.5, display=False, debug=False, aug=None):
        """
        Display the current state of the generator
        :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:
@@ -90,6 +108,12 @@

            # Scale & rotate card image
            img_card = cv2.resize(card.img, (int(len(card.img[0]) * card.scale), int(len(card.img) * card.scale)))
            if aug is not None:
                seq = iaa.Sequential([
                    iaa.SimplexNoiseAlpha(first=iaa.Add(random.randrange(128)), size_px_max=[1, 3],
                                          upscale_method="cubic"),  # Lighting
                ])
                img_card = seq.augment_image(img_card)
            mask_scale = cv2.resize(card_mask, (int(len(card_mask[0]) * card.scale), int(len(card_mask) * card.scale)))
            img_mask = cv2.bitwise_and(img_card, mask_scale)
            img_rotate = imutils.rotate_bound(img_mask, card.theta / math.pi * 180)
@@ -119,9 +143,10 @@
                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)
@@ -131,6 +156,24 @@
        '''
        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)

        if display:
            cv2.imshow('Result', img_result)
            cv2.waitKey(0)
@@ -138,7 +181,7 @@
        self.img_result = img_result
        pass

    def generate_horizontal_span(self, gap=None, scale=None, shift=None, jitter=None):
    def generate_horizontal_span(self, gap=None, scale=None, theta=0, shift=None, jitter=None):
        """
        Generating the first scenario where the cards are laid out in a straight horizontal line
        :return: True if successfully generated, otherwise False
@@ -170,10 +213,12 @@
            card.theta = 0
            card.shift(shift, shift)
            card.rotate(jitter)
            card.rotate(theta, centre=(self.width // 2 - x_anchor, self.height // 2 - y_anchor))
            x_anchor -= gap

        return True

    def generate_vertical_span(self, gap=None, scale=None, shift=None, jitter=None):
    def generate_vertical_span(self, gap=None, scale=None, theta=0, shift=None, jitter=None):
        """
        Generating the second scenario where the cards are laid out in a straight vertical line
        :return: True if successfully generated, otherwise False
@@ -206,6 +251,7 @@
            card.theta = 0
            card.shift(shift, shift)
            card.rotate(jitter)
            card.rotate(theta, centre=(self.width // 2 - x_anchor, self.height // 2 - y_anchor))
            y_anchor += gap
        return True

@@ -286,12 +332,12 @@
                #print("%s: %.1f visible" % (ext_obj.label, visible_area / obj_area * 100))
                ext_obj.visible = obj_area * visibility <= visible_area

    def export_training_data(self, out_name, visibility=0.5):
    def export_training_data(self, out_name, visibility=0.5, aug=None):
        """
        Export the generated training image along with the txt file for all bounding boxes
        :return: none
        """
        self.render(visibility)
        self.render(visibility, aug=aug)
        cv2.imwrite(out_name + '.jpg', self.img_result)
        out_txt = open(out_name+ '.txt', 'w')
        for card in self.cards:
@@ -301,7 +347,8 @@
                coords_in_gen = [card.coordinate_in_generator(key_pt[0], key_pt[1]) for key_pt in ext_obj.key_pts]
                obj_yolo_info = key_pts_to_yolo(coords_in_gen, self.width, self.height)
                if ext_obj.label == 'card':
                    out_txt.write('0 %.6f %.6f %.6f %.6f\n' % obj_yolo_info)
                    class_id = self.class_ids[card.info['name']]
                    out_txt.write(str(class_id) + ' %.6f %.6f %.6f %.6f\n' % obj_yolo_info)
                    pass
                elif ext_obj.label[:ext_obj.label.find[':']] == 'mana_symbol':
                    # TODO
@@ -446,27 +493,41 @@

def main():
    random.seed()
    ia.seed(random.randrange(10000))

    bg_images = generate_data.load_dtd(dump_it=False)
    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()
    for set_name in fetch_data.all_set_list:
        df = fetch_data.load_all_cards_text('data/csv/%s.csv' % set_name)
        card_pool = card_pool.append(df)

    num_gen = 25600
    num_iter = 3
    #card_pool = pd.DataFrame()
    #for set_name in fetch_data.all_set_list:
    #    df = fetch_data.load_all_cards_text('%s/csv/%s.csv' % (data_dir, set_name))
    #    card_pool = card_pool.append(df)
    card_pool = fetch_data.load_all_cards_text('%s/csv/custom.csv' % data_dir)
    class_ids = {}
    with open('%s/obj.names' % data_dir) as names_file:
        class_name_list = names_file.read().splitlines()
        for i in range(len(class_name_list)):
            class_ids[class_name_list[i]] = i
    print(class_ids)

    num_gen = 60000
    num_iter = 1

    for i in range(num_gen):
        generator = ImageGenerator(background.get_random(), 1440, 960)
        out_name = 'data/train/non_obstructive/'
        # 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(), class_ids, 1440, 960, skew=skew)
        out_name = ''
        for _, card_info in card_pool.sample(random.randint(2, 5)).iterrows():
            img_name = '../usb/data/png/%s/%s_%s.png' % (card_info['set'], card_info['collector_number'],
            img_name = '%s/card_img/png/%s/%s_%s.png' % (data_dir, card_info['set'], card_info['collector_number'],
                                                         fetch_data.get_valid_filename(card_info['name']))
            out_name += '%s%s_' % (card_info['set'], card_info['collector_number'])
            card_img = cv2.imread(img_name)
            if card_img is None:
                fetch_data.fetch_card_image(card_info, out_dir='../usb/data/png/%s' % card_info['set'])
                fetch_data.fetch_card_image(card_info, out_dir='%s/card_img/png/%s' % (data_dir, card_info['set']))
                card_img = cv2.imread(img_name)
            if card_img is None:
                print('WARNING: card %s is not found!' % img_name)
@@ -474,51 +535,31 @@
            card = Card(card_img, card_info, detected_object_list)
            generator.add_card(card)
        for j in range(num_iter):
            generator.generate_non_obstructive()
            #generator.generate_horizontal_span()
            generator.export_training_data(visibility=0.0, out_name=out_name + str(j))
            seq = iaa.Sequential([
                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, 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()
                generator.export_training_data(visibility=0.0, out_name='%s/train/non_obstructive_10/%s%d'
                                                                        % (data_dir, out_name, j), aug=seq)
            elif i % 3 == 1:
                generator.generate_horizontal_span(theta=random.uniform(-math.pi, math.pi))
                generator.export_training_data(visibility=0.0, out_name='%s/train/horizontal_span_10/%s%d'
                                                                        % (data_dir, out_name, j), aug=seq)
            else:
                generator.generate_vertical_span(theta=random.uniform(-math.pi, math.pi))
                generator.export_training_data(visibility=0.0, out_name='%s/train/vertical_span_10/%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()

    '''
    #img_bg = cv2.imread('data/frilly_0007.jpg')
    #generator = ImageGenerator(img_bg, 1440, 960)
    card_pool = pd.DataFrame()
    for set_name in fetch_data.all_set_list:
        df = fetch_data.load_all_cards_text('data/csv/%s.csv' % set_name)
        card_info = df.iloc[random.randint(0, df.shape[0] - 1)]
        # Currently ignoring planeswalker cards due to their different card layout
        is_planeswalker = 'Planeswalker' in card_info['type_line']
        if not is_planeswalker:
            card_pool = card_pool.append(card_info)
    for i in [random.randrange(0, card_pool.shape[0] - 1, 1) for _ in range(4)]:
        card_info = card_pool.iloc[i]
        img_name = '../usb/data/png/%s/%s_%s.png' % (card_info['set'], card_info['collector_number'],
                                                     fetch_data.get_valid_filename(card_info['name']))
        print(img_name)
        card_img = cv2.imread(img_name)
        if card_img is None:
            fetch_data.fetch_card_image(card_info, out_dir='../usb/data/png/%s' % card_info['set'])
            card_img = cv2.imread(img_name)
        detected_object_list = generate_data.apply_bounding_box(card_img, card_info)
        card = Card(card_img, card_info, detected_object_list)

        generator.add_card(card)
        #generator.add_card(card, x=random.uniform(200, generator.width - 200),
        #                   y=random.uniform(200, generator.height - 200), theta=random.uniform(-math.pi, math.pi), scale=0.5)
        #card.shift([-100, 100], [-100, 100])
        #card.rotate((0, 0), [-math.pi / 4, math.pi / 4])
    import time

    for i in range(100):
        generator.generate_vertical_span()
        generator.render(debug=False)
        generator.export_training_data(out_name='data/test')
    #generator.generate_horizontal_span()
    #generator.render(debug=True)
    #generator.generate_vertical_span()
    #generator.render(debug=True)
    '''
    pass