xref: /OK3568_Linux_fs/external/rknn-toolkit2/examples/darknet/yolov3_416x416/yolov3_utils.py (revision 4882a59341e53eb6f0b4789bf948001014eff981)
1import numpy as np
2import cv2
3import os
4import urllib.request
5
6NUM_CLS = 80
7MAX_BOXES = 500
8OBJ_THRESH = 0.5
9NMS_THRESH = 0.6
10
11CLASSES = ("person", "bicycle", "car","motorbike ","aeroplane ","bus ","train","truck ","boat","traffic light",
12           "fire hydrant","stop sign ","parking meter","bench","bird","cat","dog ","horse ","sheep","cow","elephant",
13           "bear","zebra ","giraffe","backpack","umbrella","handbag","tie","suitcase","frisbee","skis","snowboard","sports ball","kite",
14           "baseball bat","baseball glove","skateboard","surfboard","tennis racket","bottle","wine glass","cup","fork","knife ",
15           "spoon","bowl","banana","apple","sandwich","orange","broccoli","carrot","hot dog","pizza ","donut","cake","chair","sofa",
16           "pottedplant","bed","diningtable","toilet ","tvmonitor","laptop	","mouse	","remote ","keyboard ","cell phone","microwave ",
17           "oven ","toaster","sink","refrigerator ","book","clock","vase","scissors ","teddy bear ","hair drier", "toothbrush ")
18
19def sigmoid(x):
20    return 1 / (1 + np.exp(-x))
21
22def process(input, mask, anchors):
23
24    anchors = [anchors[i] for i in mask]
25    grid_h, grid_w = map(int, input.shape[0:2])
26
27    box_confidence = sigmoid(input[..., 4])
28    box_confidence = np.expand_dims(box_confidence, axis=-1)
29
30    box_class_probs = sigmoid(input[..., 5:])
31
32    box_xy = sigmoid(input[..., :2])
33    box_wh = np.exp(input[..., 2:4])
34    box_wh = box_wh * anchors
35
36    col = np.tile(np.arange(0, grid_w), grid_w).reshape(-1, grid_w)
37    row = np.tile(np.arange(0, grid_h).reshape(-1, 1), grid_h)
38
39    col = col.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2)
40    row = row.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2)
41    grid = np.concatenate((col, row), axis=-1)
42
43    box_xy += grid
44    box_xy /= (grid_w, grid_h)
45    box_wh /= (416, 416)
46    box_xy -= (box_wh / 2.)
47    box = np.concatenate((box_xy, box_wh), axis=-1)
48
49    return box, box_confidence, box_class_probs
50
51def filter_boxes(boxes, box_confidences, box_class_probs):
52    """Filter boxes with object threshold.
53
54    # Arguments
55        boxes: ndarray, boxes of objects.
56        box_confidences: ndarray, confidences of objects.
57        box_class_probs: ndarray, class_probs of objects.
58
59    # Returns
60        boxes: ndarray, filtered boxes.
61        classes: ndarray, classes for boxes.
62        scores: ndarray, scores for boxes.
63    """
64    box_scores = box_confidences * box_class_probs
65    box_classes = np.argmax(box_scores, axis=-1)
66    box_class_scores = np.max(box_scores, axis=-1)
67    pos = np.where(box_class_scores >= OBJ_THRESH)
68
69    boxes = boxes[pos]
70    classes = box_classes[pos]
71    scores = box_class_scores[pos]
72
73    return boxes, classes, scores
74
75def nms_boxes(boxes, scores):
76    """Suppress non-maximal boxes.
77
78    # Arguments
79        boxes: ndarray, boxes of objects.
80        scores: ndarray, scores of objects.
81
82    # Returns
83        keep: ndarray, index of effective boxes.
84    """
85    x = boxes[:, 0]
86    y = boxes[:, 1]
87    w = boxes[:, 2]
88    h = boxes[:, 3]
89
90    areas = w * h
91    order = scores.argsort()[::-1]
92
93    keep = []
94    while order.size > 0:
95        i = order[0]
96        keep.append(i)
97
98        xx1 = np.maximum(x[i], x[order[1:]])
99        yy1 = np.maximum(y[i], y[order[1:]])
100        xx2 = np.minimum(x[i] + w[i], x[order[1:]] + w[order[1:]])
101        yy2 = np.minimum(y[i] + h[i], y[order[1:]] + h[order[1:]])
102
103        w1 = np.maximum(0.0, xx2 - xx1 + 0.00001)
104        h1 = np.maximum(0.0, yy2 - yy1 + 0.00001)
105        inter = w1 * h1
106
107        ovr = inter / (areas[i] + areas[order[1:]] - inter)
108        inds = np.where(ovr <= NMS_THRESH)[0]
109        order = order[inds + 1]
110    keep = np.array(keep)
111    return keep
112
113def yolov3_post_process(input_data):
114    # yolov3
115    masks = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
116    anchors = [[10, 13], [16, 30], [33, 23], [30, 61], [62, 45],
117              [59, 119], [116, 90], [156, 198], [373, 326]]
118    # yolov3-tiny
119    # masks = [[3, 4, 5], [0, 1, 2]]
120    # anchors = [[10, 14], [23, 27], [37, 58], [81, 82], [135, 169], [344, 319]]
121
122    boxes, classes, scores = [], [], []
123    for input,mask in zip(input_data, masks):
124        b, c, s = process(input, mask, anchors)
125        b, c, s = filter_boxes(b, c, s)
126        boxes.append(b)
127        classes.append(c)
128        scores.append(s)
129
130    boxes = np.concatenate(boxes)
131    classes = np.concatenate(classes)
132    scores = np.concatenate(scores)
133
134    nboxes, nclasses, nscores = [], [], []
135    for c in set(classes):
136        inds = np.where(classes == c)
137        b = boxes[inds]
138        c = classes[inds]
139        s = scores[inds]
140
141        keep = nms_boxes(b, s)
142
143        nboxes.append(b[keep])
144        nclasses.append(c[keep])
145        nscores.append(s[keep])
146
147    if not nclasses and not nscores:
148        return None, None, None
149
150    boxes = np.concatenate(nboxes)
151    classes = np.concatenate(nclasses)
152    scores = np.concatenate(nscores)
153
154    return boxes, classes, scores
155
156def draw(image, boxes, scores, classes):
157    """Draw the boxes on the image.
158
159    # Argument:
160        image: original image.
161        boxes: ndarray, boxes of objects.
162        classes: ndarray, classes of objects.
163        scores: ndarray, scores of objects.
164        all_classes: all classes name.
165    """
166    for box, score, cl in zip(boxes, scores, classes):
167        x, y, w, h = box
168        print('class: {}, score: {}'.format(CLASSES[cl], score))
169        print('box coordinate left,top,right,down: [{}, {}, {}, {}]'.format(x, y, x+w, y+h))
170        x *= image.shape[1]
171        y *= image.shape[0]
172        w *= image.shape[1]
173        h *= image.shape[0]
174        top = max(0, np.floor(x + 0.5).astype(int))
175        left = max(0, np.floor(y + 0.5).astype(int))
176        right = min(image.shape[1], np.floor(x + w + 0.5).astype(int))
177        bottom = min(image.shape[0], np.floor(y + h + 0.5).astype(int))
178
179        cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 2)
180        cv2.putText(image, '{0} {1:.2f}'.format(CLASSES[cl], score),
181                    (top, left - 6),
182                    cv2.FONT_HERSHEY_SIMPLEX,
183                    0.6, (0, 0, 255), 2)
184
185
186def download_yolov3_weight(dst_path):
187    if os.path.exists(dst_path):
188        print('yolov3.weight exist.')
189        return
190    print('Downloading yolov3.weights...')
191    url = 'https://pjreddie.com/media/files/yolov3.weights'
192    try:
193        urllib.request.urlretrieve(url, dst_path)
194    except urllib.error.HTTPError as e:
195        print('HTTPError code: ', e.code)
196        print('HTTPError reason: ', e.reason)
197        exit(-1)
198    except urllib.error.URLError as e:
199        print('URLError reason: ', e.reason)
200    else:
201        print('Download yolov3.weight success.')
202