YOLOv3での検出範囲選択(マウスイベント)

前提・実現したいこと

YOLOv3(keras版)を改良し，マウスイベントを用いて物体の検出範囲が指定できるようにしたいです。
そこで，下記リンクのソースコードを参考にyolo.pyに書き足す形で考えたのですが，どの辺りを変更していいのかがわかりません。

マウスイベント参考先↓
https://qiita.com/otakoma/items/04e525ac74b7191dffe6

リアルタイムでの実装を想定しているため
動画読み込む→物体検出→マウスでエリア(台形)を指定→検出された物体のエリア内外判定
という流れで動作することが理想です。

※ご指摘頂いた通り、この流れだとマウスでエリアを指定している間は読み込んだ動画が流れていってしまいます。リアルタイムでの長時間の検出が目的な為、動画が流れてしまっても大丈夫です。

検出した物体のエリア内外判定についてはopen cvの関数「PointPolygonTest」を用いて実装する予定です

該当のソースコード

Python
1# -*- coding: utf-8 -*-
2"""
3Class definition of YOLO_v3 style detection model on image and video
4"""
5
6import colorsys
7import os
8from timeit import default_timer as timer
9import cv2
10
11import numpy as np
12from keras import backend as K
13from keras.models import load_model
14from keras.layers import Input
15from PIL import Image, ImageFont, ImageDraw
16from tensorflow.python.ops.gen_array_ops import fill
17
18from yolo3.model import yolo_eval, yolo_body, tiny_yolo_body
19from yolo3.utils import letterbox_image
20import os
21from keras.utils import multi_gpu_model
22
23class YOLO(object):
24    _defaults = {
25        "model_path": 'model_data/yolo.h5',
26        "anchors_path": 'model_data/yolo_anchors.txt',
27        "classes_path": 'model_data/coco_classes.txt',
28        "score" : 0.3,
29        "iou" : 0.45,
30        "model_image_size" : (416, 416),
31        "gpu_num" : 1,
32    }
33
34    @classmethod
35    def get_defaults(cls, n):
36        if n in cls._defaults:
37            return cls._defaults[n]
38        else:
39            return "Unrecognized attribute name '" + n + "'"
40
41    def __init__(self, **kwargs):
42        self.__dict__.update(self._defaults) # set up default values
43        self.__dict__.update(kwargs) # and update with user overrides
44        self.class_names = self._get_class()
45        self.anchors = self._get_anchors()
46        self.sess = K.get_session()
47        self.boxes, self.scores, self.classes = self.generate()
48
49    def _get_class(self):
50        classes_path = os.path.expanduser(self.classes_path)
51        with open(classes_path) as f:
52            class_names = f.readlines()
53        class_names = [c.strip() for c in class_names]
54        return class_names
55
56    def _get_anchors(self):
57        anchors_path = os.path.expanduser(self.anchors_path)
58        with open(anchors_path) as f:
59            anchors = f.readline()
60        anchors = [float(x) for x in anchors.split(',')]
61        return np.array(anchors).reshape(-1, 2)
62
63    def generate(self):
64        model_path = os.path.expanduser(self.model_path)
65        assert model_path.endswith('.h5'), 'Keras model or weights must be a .h5 file.'
66
67        # Load model, or construct model and load weights.
68        num_anchors = len(self.anchors)
69        num_classes = len(self.class_names)
70        is_tiny_version = num_anchors==6 # default setting
71        try:
72            self.yolo_model = load_model(model_path, compile=False)
73        except:
74            self.yolo_model = tiny_yolo_body(Input(shape=(None,None,3)), num_anchors//2, num_classes) \
75                if is_tiny_version else yolo_body(Input(shape=(None,None,3)), num_anchors//3, num_classes)
76            self.yolo_model.load_weights(self.model_path) # make sure model, anchors and classes match
77        else:
78            assert self.yolo_model.layers[-1].output_shape[-1] == \
79                num_anchors/len(self.yolo_model.output) * (num_classes + 5), \
80                'Mismatch between model and given anchor and class sizes'
81
82        print('{} model, anchors, and classes loaded.'.format(model_path))
83
84        # Generate colors for drawing bounding boxes.
85        hsv_tuples = [(x / len(self.class_names), 1., 1.)
86                      for x in range(len(self.class_names))]
87        self.colors = list(map(lambda x: colorsys.hsv_to_rgb(*x), hsv_tuples))
88        self.colors = list(
89            map(lambda x: (int(x[0] * 255), int(x[1] * 255), int(x[2] * 255)),
90                self.colors))
91        np.random.seed(10101)  # Fixed seed for consistent colors across runs.
92        np.random.shuffle(self.colors)  # Shuffle colors to decorrelate adjacent classes.
93        np.random.seed(None)  # Reset seed to default.
94
95        # Generate output tensor targets for filtered bounding boxes.
96        self.input_image_shape = K.placeholder(shape=(2, ))
97        if self.gpu_num>=2:
98            self.yolo_model = multi_gpu_model(self.yolo_model, gpus=self.gpu_num)
99        boxes, scores, classes = yolo_eval(self.yolo_model.output, self.anchors,
100                len(self.class_names), self.input_image_shape,
101                score_threshold=self.score, iou_threshold=self.iou)
102        return boxes, scores, classes
103
104    
105    def detect_image(self, image):
106        import cv2
107        start = timer()
108
109        if self.model_image_size != (None, None):
110            assert self.model_image_size[0]%32 == 0, 'Multiples of 32 required'
111            assert self.model_image_size[1]%32 == 0, 'Multiples of 32 required'
112            boxed_image = letterbox_image(image, tuple(reversed(self.model_image_size)))
113        else:
114            new_image_size = (image.width - (image.width % 32),
115                              image.height - (image.height % 32))
116            boxed_image = letterbox_image(image, new_image_size)
117        image_data = np.array(boxed_image, dtype='float32')
118
119        print(image_data.shape)
120        image_data /= 255.
121        image_data = np.expand_dims(image_data, 0)  # Add batch dimension.
122
123        out_boxes, out_scores, out_classes = self.sess.run(
124            [self.boxes, self.scores, self.classes],
125            feed_dict={
126                self.yolo_model.input: image_data,
127                self.input_image_shape: [image.size[1], image.size[0]],
128                K.learning_phase(): 0
129            })
130
131        print('Found {} boxes for {}'.format(len(out_boxes), 'img'))
132
133        font = ImageFont.truetype(font='font/FiraMono-Medium.otf',
134                    size=np.floor(3e-2 * image.size[1] + 0.5).astype('int32'))
135        thickness = (image.size[0] + image.size[1]) // 300
136
137        for i, c in reversed(list(enumerate(out_classes))):
138            predicted_class = self.class_names[c]
139            if predicted_class == 'person':
140                box = out_boxes[i]
141                score = out_scores[i]
142
143                label = '{} {:.2f}'.format(predicted_class, score)
144                draw = ImageDraw.Draw(image)
145                label_size = draw.textsize(label, font)
146
147                top, left, bottom, right = box
148                top = max(0, np.floor(top + 0.5).astype('int32'))
149                left = max(0, np.floor(left + 0.5).astype('int32'))
150                bottom = min(image.size[1], np.floor(bottom + 0.5).astype('int32'))
151                right = min(image.size[0], np.floor(right + 0.5).astype('int32'))
152                print(label, (left, top), (right, bottom))
153
154                if top - label_size[1] >= 0:
155                    text_origin = np.array([left, top - label_size[1]])
156                else:
157                    text_origin = np.array([left, top + 1])
158
159                # My kingdom for a good redistributable image drawing library.
160                for i in range(thickness):
161                    draw.rectangle(
162                        [left, top, right, bottom],
163                        outline=self.colors[c])
164                draw.rectangle(
165                    [tuple(text_origin), tuple(text_origin + label_size)],
166                    fill=self.colors[c])
167                draw.text(text_origin, label, fill=(0, 0, 0), font=font)
168                del draw
169
170        end = timer()
171        print(end - start)
172        return image
173
174    def close_session(self):
175        self.sess.close()
176
177def detect_video(yolo, video_path, output_path=""):
178    vid = cv2.VideoCapture(video_path)
179    if not vid.isOpened():
180        raise IOError("Couldn't open webcam or video")
181    video_FourCC    = int(vid.get(cv2.CAP_PROP_FOURCC))
182    video_fps       = vid.get(cv2.CAP_PROP_FPS)
183    video_size      = (int(vid.get(cv2.CAP_PROP_FRAME_WIDTH)),
184                        int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT)))
185    isOutput = True if output_path != "" else False
186    if isOutput:
187        print("!!! TYPE:", type(output_path), type(video_FourCC), type(video_fps), type(video_size))
188        out = cv2.VideoWriter(output_path, video_FourCC, video_fps, video_size)
189    accum_time = 0
190    curr_fps = 0
191    fps = "FPS: ??"
192    prev_time = timer()
193    while True:
194        return_value, frame = vid.read()
195        image = Image.fromarray(frame)
196        image = yolo.detect_image(image)
197        result = np.asarray(image)
198        curr_time = timer()
199        exec_time = curr_time - prev_time
200        prev_time = curr_time
201        accum_time = accum_time + exec_time
202        curr_fps = curr_fps + 1
203        if accum_time > 1:
204            accum_time = accum_time - 1
205            fps = "FPS:" + str(curr_fps)
206            curr_fps = 0
207        cv2.putText(result, text=fps, org=(3, 15), fontFace=cv2.FONT_HERSHEY_SIMPLEX,
208                    fontScale=0.30, color=(255, 0, 0), thickness=1)
209        cv2.namedWindow("result", cv2.WINDOW_NORMAL)
210        cv2.imshow("result", result)
211        if isOutput:
212            out.write(result)
213        if cv2.waitKey(1) & 0xFF == ord('q'):
214            break
215    yolo.close_session()
216

補足情報

yolov3 - https://github.com/qqwweee/keras-yolo3

環境
・Anaconda / Python 3.7
・keras 2.2.4
・tensorflow 1.13.1
・matplotlib
・pillow
・OpenCV

退会済みユーザー

2021/12/16 21:22

YOLOの推論結果をベースに、アノテーション結果を修正するようなツールを作りたいという意味でしょうか？ご存知と思いますが、動画はパラパラ漫画と同じでリアルタイムに画像を差し替えていく処理をしています。マウスで領域選択をする間も画像が差し替えられていくと思うのでこの辺りをどうしたいか明記した方がいいと思います。