This repository has been archived by the owner on Sep 11, 2019. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 2
/
model_box.py
256 lines (197 loc) · 8.38 KB
/
model_box.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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
# Copyright 2019 Amazon.com, Inc. or its affiliates. All Rights Reserved.
# SPDX-License-Identifier: Apache-2.0
# -*- coding: utf-8 -*-
# File: model_box.py
import numpy as np
from collections import namedtuple
import tensorflow as tf
from tensorpack.tfutils.scope_utils import under_name_scope
from config import config
@under_name_scope()
def clip_boxes_batch(boxes, window, batch_ids, name=None):
"""
Args:
boxes: nx(#class)x4, xyxy
window: BSx2
batch_ids: 1-D Tensor of size n
"""
boxes = tf.maximum(boxes, 0.0)
wh = tf.reverse(tf.gather(window, batch_ids), axis=[1])
whwh = tf.concat((wh, wh), axis=1)
multiples = tf.stack([tf.constant(1), tf.shape(boxes)[1], tf.constant(1)])
whwh_tiled = tf.tile(tf.expand_dims(whwh, 1), multiples) # asd = [1, #class, 1]
boxes = tf.minimum(boxes, tf.cast(whwh_tiled, tf.float32), name=name)
return boxes
@under_name_scope()
def clip_boxes(boxes, window, name=None):
"""
Args:
boxes: nx4, xyxy
window: [h, w]
"""
window = tf.squeeze(window, axis=0)
boxes = tf.maximum(boxes, 0.0)
m = tf.tile(tf.reverse(window, [0]), [2]) # (4,)
boxes = tf.minimum(boxes, tf.cast(m, tf.float32), name=name)
return boxes
@under_name_scope()
def decode_bbox_target(box_predictions, anchors):
"""
Args:
box_predictions: (..., 4), logits
anchors: (..., 4), floatbox. Must have the same shape
Returns:
box_decoded: (..., 4), float32. With the same shape.
"""
orig_shape = tf.shape(anchors)
box_pred_txtytwth = tf.reshape(box_predictions, (-1, 2, 2))
box_pred_txty, box_pred_twth = tf.split(box_pred_txtytwth, 2, axis=1)
# each is (...)x1x2
anchors_x1y1x2y2 = tf.reshape(anchors, (-1, 2, 2))
anchors_x1y1, anchors_x2y2 = tf.split(anchors_x1y1x2y2, 2, axis=1)
waha = anchors_x2y2 - anchors_x1y1
xaya = (anchors_x2y2 + anchors_x1y1) * 0.5
clip = np.log(config.PREPROC.MAX_SIZE / 16.)
wbhb = tf.exp(tf.minimum(box_pred_twth, clip)) * waha
xbyb = box_pred_txty * waha + xaya
x1y1 = xbyb - wbhb * 0.5
x2y2 = xbyb + wbhb * 0.5 # (...)x1x2
out = tf.concat([x1y1, x2y2], axis=-2)
return tf.reshape(out, orig_shape)
@under_name_scope()
def encode_bbox_target(boxes, anchors):
"""
Args:
boxes: (..., 4), float32
anchors: (..., 4), float32
Returns:
box_encoded: (..., 4), float32 with the same shape.
"""
anchors_x1y1x2y2 = tf.reshape(anchors, (-1, 2, 2))
anchors_x1y1, anchors_x2y2 = tf.split(anchors_x1y1x2y2, 2, axis=1)
waha = anchors_x2y2 - anchors_x1y1
xaya = (anchors_x2y2 + anchors_x1y1) * 0.5
boxes_x1y1x2y2 = tf.reshape(boxes, (-1, 2, 2))
boxes_x1y1, boxes_x2y2 = tf.split(boxes_x1y1x2y2, 2, axis=1)
wbhb = boxes_x2y2 - boxes_x1y1
xbyb = (boxes_x2y2 + boxes_x1y1) * 0.5
# Note that here not all boxes are valid. Some may be zero
txty = (xbyb - xaya) / waha
twth = tf.log(wbhb / waha) # may contain -inf for invalid boxes
encoded = tf.concat([txty, twth], axis=1) # (-1x2x2)
return tf.reshape(encoded, tf.shape(boxes))
@under_name_scope()
def crop_and_resize(image, boxes, box_ind, crop_size, orig_image_dims, pad_border=True, verbose_batch_index=None):
"""
Aligned version of tf.image.crop_and_resize, following our definition of floating point boxes.
Args:
image: NCHW
boxes: nx4, x1y1x2y2
box_ind: (n,)
crop_size (int):
Returns:
n,C,size,size
"""
prefix = "custom crop_and_resize"
if verbose_batch_index is not None:
prefix += f' B_IDX {verbose_batch_index}'
assert isinstance(crop_size, int), crop_size
# boxes = print_runtime_shape("boxes (early)", boxes, prefix=prefix)
boxes = tf.stop_gradient(boxes)
# org_img_size = tf.concat((tf.constant([-1, -1], dtype=tf.int32), orig_image_dims), axis=0)
# image = tf.slice(image, begin=tf.zeros(4, dtype=tf.int32), size=org_img_size)
image = image[:, :, :orig_image_dims[0], :orig_image_dims[1]]
# TF's crop_and_resize produces zeros on border
if pad_border:
# this can be quite slow
image = tf.pad(image, [[0, 0], [0, 0], [1, 1], [1, 1]], mode='SYMMETRIC')
boxes = boxes + 1
@under_name_scope()
def transform_fpcoor_for_tf(boxes, image_shape, crop_shape):
"""
The way tf.image.crop_and_resize works (with normalized box):
Initial point (the value of output[0]): x0_box * (W_img - 1)
Spacing: w_box * (W_img - 1) / (W_crop - 1)
Use the above grid to bilinear sample.
However, what we want is (with fpcoor box):
Spacing: w_box / W_crop
Initial point: x0_box + spacing/2 - 0.5
(-0.5 because bilinear sample (in my definition) assumes floating point coordinate
(0.0, 0.0) is the same as pixel value (0, 0))
This function transform fpcoor boxes to a format to be used by tf.image.crop_and_resize
Returns:
y1x1y2x2
"""
x0, y0, x1, y1 = tf.split(boxes, 4, axis=1)
spacing_w = (x1 - x0) / tf.cast(crop_shape[1], tf.float32)
spacing_h = (y1 - y0) / tf.cast(crop_shape[0], tf.float32)
imshape = [tf.cast(image_shape[0] - 1, tf.float32), tf.cast(image_shape[1] - 1, tf.float32)]
nx0 = (x0 + spacing_w / 2 - 0.5) / imshape[1]
ny0 = (y0 + spacing_h / 2 - 0.5) / imshape[0]
nw = spacing_w * tf.cast(crop_shape[1] - 1, tf.float32) / imshape[1]
nh = spacing_h * tf.cast(crop_shape[0] - 1, tf.float32) / imshape[0]
return tf.concat([ny0, nx0, ny0 + nh, nx0 + nw], axis=1)
# Expand bbox to a minium size of 1
# boxes_x1y1, boxes_x2y2 = tf.split(boxes, 2, axis=1)
# boxes_wh = boxes_x2y2 - boxes_x1y1
# boxes_center = tf.reshape((boxes_x2y2 + boxes_x1y1) * 0.5, [-1, 2])
# boxes_newwh = tf.maximum(boxes_wh, 1.)
# boxes_x1y1new = boxes_center - boxes_newwh * 0.5
# boxes_x2y2new = boxes_center + boxes_newwh * 0.5
# boxes = tf.concat([boxes_x1y1new, boxes_x2y2new], axis=1)
boxes = transform_fpcoor_for_tf(boxes, orig_image_dims, [crop_size, crop_size])
image = tf.transpose(image, [0, 2, 3, 1]) # nhwc
# image = print_runtime_shape("image", image, prefix=prefix)
# boxes = print_runtime_shape("boxes (late)", boxes, prefix=prefix)
ret = tf.image.crop_and_resize(
image, boxes, tf.cast(box_ind, tf.int32),
crop_size=[crop_size, crop_size])
ret = tf.transpose(ret, [0, 3, 1, 2]) # ncss
return ret
class RPNAnchors(namedtuple('_RPNAnchors', ['boxes', 'gt_labels', 'gt_boxes'])):
"""
boxes (FS x FS x NA x 4): The anchor boxes.
gt_labels (FS x FS x NA):
gt_boxes (FS x FS x NA x 4): Groundtruth boxes corresponding to each anchor.
"""
def encoded_gt_boxes(self):
return encode_bbox_target(self.gt_boxes, self.boxes)
def decode_logits(self, logits):
return decode_bbox_target(logits, self.boxes)
@under_name_scope()
def narrow_to_featuremap_dims(self, featuremap_dims):
"""
Take in h and w of featuremap that we need to narrow to.
Separate batch implementation because we need to handle padding.
# Boxes don't have a batch_dim when this function is called - and neither should
"""
slice3d = tf.concat([featuremap_dims, [-1]], axis=0)
slice4d = tf.concat([featuremap_dims, [-1, -1]], axis=0)
boxes = tf.slice(self.boxes, [0, 0, 0, 0], slice4d)
gt_labels = tf.slice(self.gt_labels, [0, 0, 0], slice3d)
gt_boxes = tf.slice(self.gt_boxes, [0, 0, 0, 0], slice4d)
return RPNAnchors(boxes, gt_labels, gt_boxes)
if __name__ == '__main__':
"""
Demonstrate what's wrong with tf.image.crop_and_resize:
"""
import tensorflow.contrib.eager as tfe
tfe.enable_eager_execution()
# want to crop 2x2 out of a 5x5 image, and resize to 4x4
image = np.arange(25).astype('float32').reshape(5, 5)
boxes = np.asarray([[1, 1, 3, 3]], dtype='float32')
target = 4
print(crop_and_resize(
image[None, None, :, :], boxes, [0], target)[0][0])
"""
Expected values:
4.5 5 5.5 6
7 7.5 8 8.5
9.5 10 10.5 11
12 12.5 13 13.5
You cannot easily get the above results with tf.image.crop_and_resize.
Try out yourself here:
"""
print(tf.image.crop_and_resize(
image[None, :, :, None],
np.asarray([[1, 1, 2, 2]]) / 4.0, [0], [target, target])[0][:, :, 0])