forked from rykov8/ssd_keras
-
Notifications
You must be signed in to change notification settings - Fork 85
/
tbpp_layers.py
216 lines (155 loc) · 8.93 KB
/
tbpp_layers.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
import numpy as np
import tensorflow as tf
import tensorflow_addons as tfa
import keras.backend as K
from keras.layers import Layer
class TBPPDecodeAndCrop(Layer):
"""Layer for decoding TextBoxes++ output and cropping the detected text instances from the input image.
# Arguments
prior_util: instance of TextBoxes++ PriorUtility
confidence_threshold, iou_threshold, top_k: same arguments as those that would be passed to the decoder method of PriorUtility
output_size: tuple, output size of the padded text instances, (32, 256) by default
# Notes
requires tensorflow-addons
does not work with changing image size
limited to one class
provieds grayscale images to the recogition stage
"""
def __init__(self, prior_util, confidence_threshold=0.01, iou_threshold=0.45, top_k=200, output_size=(32, 256), **kwargs):
self.prior_util = prior_util
self.confidence_threshold = confidence_threshold
self.iou_threshold = iou_threshold
self.top_k = top_k
self.output_size = output_size
self.output_height, self.output_width = output_size
super(TBPPDecodeAndCrop, self).__init__(**kwargs)
def build(self, input_shape):
image_size = self.prior_util.image_size
priors_xy = self.prior_util.priors_xy / image_size
priors_wh = self.prior_util.priors_wh / image_size
priors_variances = self.prior_util.priors_variances
self.priors_xy = tf.constant(priors_xy, dtype=K.floatx())
self.priors_wh = tf.constant(priors_wh, dtype=K.floatx())
self.priors_variances = tf.constant(priors_variances, dtype=K.floatx())
priors_minmax = np.hstack([priors_xy-priors_wh/2, priors_xy+priors_wh/2])
self.ref = tf.constant(priors_minmax[:,(0,1,2,1,2,3,0,3)], dtype=K.floatx()) # corner points
super(TBPPDecodeAndCrop, self).build(input_shape)
def call(self, x):
# calculation is done with normalized sizes
# x[0] image
# x[1] local predictions
# 4 mbox_loc + 8 mbox_quad + 5 mbox_rbox + 2 mbox_conf
def for_each_sample(x):
## decoding and NMS
y_pred = x[1]
mask = y_pred[:,18] > self.confidence_threshold
boxes_to_process = tf.boolean_mask(y_pred, mask)
#priors = tf.boolean_mask(self.priors, mask)
priors_xy = tf.boolean_mask(self.priors_xy, mask)
priors_wh = tf.boolean_mask(self.priors_wh, mask)
variances = tf.boolean_mask(self.priors_variances, mask)
variances_xy = variances[:,0:2]
variances_wh = variances[:,2:4]
offsets = boxes_to_process[:,:4]
boxes_xy = priors_xy + offsets[:,0:2] * variances_xy * priors_wh
boxes_wh = priors_wh * tf.exp(offsets[:,2:4] * variances_wh)
boxes_xy_min = boxes_xy - boxes_wh / 2.
boxes_xy_max = boxes_xy + boxes_wh / 2.
boxes_minmax = tf.concat((boxes_xy_min, boxes_xy_max), axis=-1)
#boxes_minmax = tf.clip_by_value(boxes_minmax, 0.0, 1.0)
scores = boxes_to_process[:,18]
idxs = tf.image.non_max_suppression(boxes_minmax, scores,
max_output_size=self.top_k,
iou_threshold=self.iou_threshold)
good_boxes = tf.gather(boxes_to_process, idxs, axis=0)
good_minmax = tf.gather(boxes_minmax, idxs, axis=0)
ref = tf.boolean_mask(self.ref, mask)
ref = tf.gather(ref, idxs, axis=0)
priors_xy = tf.gather(priors_xy, idxs, axis=0)
priors_wh = tf.gather(priors_wh, idxs, axis=0)
variances_xy = tf.gather(variances_xy, idxs, axis=0)
variances_wh = tf.gather(variances_wh, idxs, axis=0)
offsets_quads = good_boxes[:,4:12]
good_quads = ref + offsets_quads * tf.tile(priors_wh * variances_xy, (1,4))
offsets_rboxs = good_boxes[:,12:17]
good_rboxs = tf.concat((
priors_xy + offsets_rboxs[:,0:2] * priors_wh * variances_xy,
priors_xy + offsets_rboxs[:,2:4] * priors_wh * variances_xy,
tf.exp(offsets_rboxs[:,4:5] * variances_wh[:,1:2]) * priors_wh[:,1:2]
),-1)
good_confs = good_boxes[:,18:19]
# we only have one class :)
good_labels = tf.ones_like(good_confs)
# 4 boxes + 8 quad + 5 rboxes + 1 confs + 1 labels
good_boxes = tf.concat((
good_minmax,
good_quads,
good_rboxs,
good_confs,
good_labels,
), -1)
## cropping
img = x[0]
img = tf.image.rgb_to_grayscale(img)
img = tf.expand_dims(img, 0)
img_size = tf.cast(tf.shape(img)[1:3], 'float32')
polys = good_boxes[:,4:12] * tf.tile(img_size, (4,))
n = num_polys = tf.shape(polys)[0]
h, w = self.output_height, self.output_width
p = h * 0.05
d = h/2
pad_value = 0
def crop():
tl, tr, br, bl = polys[:,0:2], polys[:,2:4], polys[:,4:6], polys[:,6:8]
box_h = (tf.norm(tl-bl, axis=-1) + tf.norm(tr-br, axis=-1)) / 2
box_w = (tf.norm(tl-tr, axis=-1) + tf.norm(bl-br, axis=-1)) / 2
w_mod = tf.clip_by_value(h*box_w/box_h, 0, w)
x = tf.stack((tf.zeros_like(w_mod)+p, w_mod-p, w_mod-p, tf.zeros_like(w_mod)+p), axis=-1)
y = tf.repeat(np.array([[p,p,h-p,h-p]], dtype='float32'), n, axis=0)
u = polys[:,0::2]
v = polys[:,1::2]
ones, zeros = tf.ones((n, 4)), tf.zeros((n, 4))
A = tf.concat([
tf.stack([x, y, ones, zeros, zeros, zeros, -x*u, -y*u], axis=-1),
tf.stack([zeros, zeros, zeros, x, y, ones, -x*v, -y*v], axis=-1),
], axis=1)
b = tf.expand_dims(tf.concat((u,v), axis=-1), axis=-1)
M = tf.squeeze(tf.linalg.solve(A, b), axis=-1)
w_max = tf.reduce_max(w_mod)
w_mod = tf.cast(w_mod, 'int32')
def for_each_box(i, offset, words, offsets):
c = w_mod[i]
word = tfa.image.transform(img, M[i], interpolation='BILINEAR', output_shape=(h,w_max))
word = tf.transpose(word, (0,2,1,3))
word = tf.concat([ word[0,:c], tf.ones((d,h,1))*pad_value ], axis=0)
words = words.write(i, word)
offset = offset + tf.cast(c, 'float32') + d
offsets = offsets.write(i, offset)
return i+1, offset, words, offsets
words = tf.TensorArray(dtype=tf.float32, infer_shape=False, size=n)
offsets = tf.TensorArray(dtype=tf.float32, infer_shape=False, size=n)
_, _, words, offsets = tf.while_loop(
lambda i, *_: tf.less(i, n), for_each_box,
[0, -d/2, words, offsets])
words = words.concat()
offsets = offsets.stack()
return words, offsets
words, offsets = tf.cond( n > 0, crop,
lambda: (tf.zeros((0,32,1)), tf.zeros(0)) )
words = words[:w,:,0]
words = tf.pad(words, ((0,w-tf.shape(words)[0]), (0,0)), constant_values=pad_value)
words = tf.expand_dims(words, axis=-1)
offsets = tf.expand_dims(offsets, axis=-1)
good_boxes = tf.concat([good_boxes, offsets], axis=-1)
good_boxes = tf.pad(good_boxes, ((0,self.top_k-tf.shape(good_boxes)[0]),(0,0)), constant_values=0)
return words, good_boxes
cropped_images, boxes = tf.map_fn(for_each_sample, x, dtype=('float32', 'float32'))
return [cropped_images, boxes]
def get_config(self):
base_config = super(TBPPDecodeAndCrop, self).get_config()
base_config['prior_util'] = self.prior_util
base_config['confidence_threshold'] = self.confidence_threshold
base_config['iou_threshold'] = self.iou_threshold
base_config['top_k'] = self.top_k
base_config['output_size'] = self.output_size
return base_config