-
Notifications
You must be signed in to change notification settings - Fork 45.8k
/
faster_rcnn_meta_arch.py
1535 lines (1376 loc) · 72.1 KB
/
faster_rcnn_meta_arch.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
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Faster R-CNN meta-architecture definition.
General tensorflow implementation of Faster R-CNN detection models.
See Faster R-CNN: Ren, Shaoqing, et al.
"Faster R-CNN: Towards real-time object detection with region proposal
networks." Advances in neural information processing systems. 2015.
We allow for two modes: first_stage_only=True and first_stage_only=False. In
the former setting, all of the user facing methods (e.g., predict, postprocess,
loss) can be used as if the model consisted only of the RPN, returning class
agnostic proposals (these can be thought of as approximate detections with no
associated class information). In the latter setting, proposals are computed,
then passed through a second stage "box classifier" to yield (multi-class)
detections.
Implementations of Faster R-CNN models must define a new
FasterRCNNFeatureExtractor and override three methods: `preprocess`,
`_extract_proposal_features` (the first stage of the model), and
`_extract_box_classifier_features` (the second stage of the model). Optionally,
the `restore_fn` method can be overridden. See tests for an example.
A few important notes:
+ Batching conventions: We support batched inference and training where
all images within a batch have the same resolution. Batch sizes are determined
dynamically via the shape of the input tensors (rather than being specified
directly as, e.g., a model constructor).
A complication is that due to non-max suppression, we are not guaranteed to get
the same number of proposals from the first stage RPN (region proposal network)
for each image (though in practice, we should often get the same number of
proposals). For this reason we pad to a max number of proposals per image
within a batch. This `self.max_num_proposals` property is set to the
`first_stage_max_proposals` parameter at inference time and the
`second_stage_batch_size` at training time since we subsample the batch to
be sent through the box classifier during training.
For the second stage of the pipeline, we arrange the proposals for all images
within the batch along a single batch dimension. For example, the input to
_extract_box_classifier_features is a tensor of shape
`[total_num_proposals, crop_height, crop_width, depth]` where
total_num_proposals is batch_size * self.max_num_proposals. (And note that per
the above comment, a subset of these entries correspond to zero paddings.)
+ Coordinate representations:
Following the API (see model.DetectionModel definition), our outputs after
postprocessing operations are always normalized boxes however, internally, we
sometimes convert to absolute --- e.g. for loss computation. In particular,
anchors and proposal_boxes are both represented as absolute coordinates.
TODO: Support TPU implementations and sigmoid loss.
"""
from abc import abstractmethod
from functools import partial
import tensorflow as tf
from object_detection.anchor_generators import grid_anchor_generator
from object_detection.core import balanced_positive_negative_sampler as sampler
from object_detection.core import box_list
from object_detection.core import box_list_ops
from object_detection.core import box_predictor
from object_detection.core import losses
from object_detection.core import model
from object_detection.core import post_processing
from object_detection.core import standard_fields as fields
from object_detection.core import target_assigner
from object_detection.utils import ops
from object_detection.utils import shape_utils
slim = tf.contrib.slim
class FasterRCNNFeatureExtractor(object):
"""Faster R-CNN Feature Extractor definition."""
def __init__(self,
is_training,
first_stage_features_stride,
reuse_weights=None,
weight_decay=0.0):
"""Constructor.
Args:
is_training: A boolean indicating whether the training version of the
computation graph should be constructed.
first_stage_features_stride: Output stride of extracted RPN feature map.
reuse_weights: Whether to reuse variables. Default is None.
weight_decay: float weight decay for feature extractor (default: 0.0).
"""
self._is_training = is_training
self._first_stage_features_stride = first_stage_features_stride
self._reuse_weights = reuse_weights
self._weight_decay = weight_decay
@abstractmethod
def preprocess(self, resized_inputs):
"""Feature-extractor specific preprocessing (minus image resizing)."""
pass
def extract_proposal_features(self, preprocessed_inputs, scope):
"""Extracts first stage RPN features.
This function is responsible for extracting feature maps from preprocessed
images. These features are used by the region proposal network (RPN) to
predict proposals.
Args:
preprocessed_inputs: A [batch, height, width, channels] float tensor
representing a batch of images.
scope: A scope name.
Returns:
rpn_feature_map: A tensor with shape [batch, height, width, depth]
"""
with tf.variable_scope(scope, values=[preprocessed_inputs]):
return self._extract_proposal_features(preprocessed_inputs, scope)
@abstractmethod
def _extract_proposal_features(self, preprocessed_inputs, scope):
"""Extracts first stage RPN features, to be overridden."""
pass
def extract_box_classifier_features(self, proposal_feature_maps, scope):
"""Extracts second stage box classifier features.
Args:
proposal_feature_maps: A 4-D float tensor with shape
[batch_size * self.max_num_proposals, crop_height, crop_width, depth]
representing the feature map cropped to each proposal.
scope: A scope name.
Returns:
proposal_classifier_features: A 4-D float tensor with shape
[batch_size * self.max_num_proposals, height, width, depth]
representing box classifier features for each proposal.
"""
with tf.variable_scope(scope, values=[proposal_feature_maps]):
return self._extract_box_classifier_features(proposal_feature_maps, scope)
@abstractmethod
def _extract_box_classifier_features(self, proposal_feature_maps, scope):
"""Extracts second stage box classifier features, to be overridden."""
pass
def restore_from_classification_checkpoint_fn(
self,
first_stage_feature_extractor_scope,
second_stage_feature_extractor_scope):
"""Returns a map of variables to load from a foreign checkpoint.
Args:
first_stage_feature_extractor_scope: A scope name for the first stage
feature extractor.
second_stage_feature_extractor_scope: A scope name for the second stage
feature extractor.
Returns:
A dict mapping variable names (to load from a checkpoint) to variables in
the model graph.
"""
variables_to_restore = {}
for variable in tf.global_variables():
for scope_name in [first_stage_feature_extractor_scope,
second_stage_feature_extractor_scope]:
if variable.op.name.startswith(scope_name):
var_name = variable.op.name.replace(scope_name + '/', '')
variables_to_restore[var_name] = variable
return variables_to_restore
class FasterRCNNMetaArch(model.DetectionModel):
"""Faster R-CNN Meta-architecture definition."""
def __init__(self,
is_training,
num_classes,
image_resizer_fn,
feature_extractor,
first_stage_only,
first_stage_anchor_generator,
first_stage_atrous_rate,
first_stage_box_predictor_arg_scope,
first_stage_box_predictor_kernel_size,
first_stage_box_predictor_depth,
first_stage_minibatch_size,
first_stage_positive_balance_fraction,
first_stage_nms_score_threshold,
first_stage_nms_iou_threshold,
first_stage_max_proposals,
first_stage_localization_loss_weight,
first_stage_objectness_loss_weight,
initial_crop_size,
maxpool_kernel_size,
maxpool_stride,
second_stage_mask_rcnn_box_predictor,
second_stage_batch_size,
second_stage_balance_fraction,
second_stage_non_max_suppression_fn,
second_stage_score_conversion_fn,
second_stage_localization_loss_weight,
second_stage_classification_loss_weight,
hard_example_miner,
parallel_iterations=16):
"""FasterRCNNMetaArch Constructor.
Args:
is_training: A boolean indicating whether the training version of the
computation graph should be constructed.
num_classes: Number of classes. Note that num_classes *does not*
include the background category, so if groundtruth labels take values
in {0, 1, .., K-1}, num_classes=K (and not K+1, even though the
assigned classification targets can range from {0,... K}).
image_resizer_fn: A callable for image resizing. This callable always
takes a rank-3 image tensor (corresponding to a single image) and
returns a rank-3 image tensor, possibly with new spatial dimensions.
See builders/image_resizer_builder.py.
feature_extractor: A FasterRCNNFeatureExtractor object.
first_stage_only: Whether to construct only the Region Proposal Network
(RPN) part of the model.
first_stage_anchor_generator: An anchor_generator.AnchorGenerator object
(note that currently we only support
grid_anchor_generator.GridAnchorGenerator objects)
first_stage_atrous_rate: A single integer indicating the atrous rate for
the single convolution op which is applied to the `rpn_features_to_crop`
tensor to obtain a tensor to be used for box prediction. Some feature
extractors optionally allow for producing feature maps computed at
denser resolutions. The atrous rate is used to compensate for the
denser feature maps by using an effectively larger receptive field.
(This should typically be set to 1).
first_stage_box_predictor_arg_scope: Slim arg_scope for conv2d,
separable_conv2d and fully_connected ops for the RPN box predictor.
first_stage_box_predictor_kernel_size: Kernel size to use for the
convolution op just prior to RPN box predictions.
first_stage_box_predictor_depth: Output depth for the convolution op
just prior to RPN box predictions.
first_stage_minibatch_size: The "batch size" to use for computing the
objectness and location loss of the region proposal network. This
"batch size" refers to the number of anchors selected as contributing
to the loss function for any given image within the image batch and is
only called "batch_size" due to terminology from the Faster R-CNN paper.
first_stage_positive_balance_fraction: Fraction of positive examples
per image for the RPN. The recommended value for Faster RCNN is 0.5.
first_stage_nms_score_threshold: Score threshold for non max suppression
for the Region Proposal Network (RPN). This value is expected to be in
[0, 1] as it is applied directly after a softmax transformation. The
recommended value for Faster R-CNN is 0.
first_stage_nms_iou_threshold: The Intersection Over Union (IOU) threshold
for performing Non-Max Suppression (NMS) on the boxes predicted by the
Region Proposal Network (RPN).
first_stage_max_proposals: Maximum number of boxes to retain after
performing Non-Max Suppression (NMS) on the boxes predicted by the
Region Proposal Network (RPN).
first_stage_localization_loss_weight: A float
first_stage_objectness_loss_weight: A float
initial_crop_size: A single integer indicating the output size
(width and height are set to be the same) of the initial bilinear
interpolation based cropping during ROI pooling.
maxpool_kernel_size: A single integer indicating the kernel size of the
max pool op on the cropped feature map during ROI pooling.
maxpool_stride: A single integer indicating the stride of the max pool
op on the cropped feature map during ROI pooling.
second_stage_mask_rcnn_box_predictor: Mask R-CNN box predictor to use for
the second stage.
second_stage_batch_size: The batch size used for computing the
classification and refined location loss of the box classifier. This
"batch size" refers to the number of proposals selected as contributing
to the loss function for any given image within the image batch and is
only called "batch_size" due to terminology from the Faster R-CNN paper.
second_stage_balance_fraction: Fraction of positive examples to use
per image for the box classifier. The recommended value for Faster RCNN
is 0.25.
second_stage_non_max_suppression_fn: batch_multiclass_non_max_suppression
callable that takes `boxes`, `scores`, optional `clip_window` and
optional (kwarg) `mask` inputs (with all other inputs already set)
and returns a dictionary containing tensors with keys:
`detection_boxes`, `detection_scores`, `detection_classes`,
`num_detections`, and (optionally) `detection_masks`. See
`post_processing.batch_multiclass_non_max_suppression` for the type and
shape of these tensors.
second_stage_score_conversion_fn: Callable elementwise nonlinearity
(that takes tensors as inputs and returns tensors). This is usually
used to convert logits to probabilities.
second_stage_localization_loss_weight: A float
second_stage_classification_loss_weight: A float
hard_example_miner: A losses.HardExampleMiner object (can be None).
parallel_iterations: (Optional) The number of iterations allowed to run
in parallel for calls to tf.map_fn.
Raises:
ValueError: If `second_stage_batch_size` > `first_stage_max_proposals`
ValueError: If first_stage_anchor_generator is not of type
grid_anchor_generator.GridAnchorGenerator.
"""
super(FasterRCNNMetaArch, self).__init__(num_classes=num_classes)
if second_stage_batch_size > first_stage_max_proposals:
raise ValueError('second_stage_batch_size should be no greater than '
'first_stage_max_proposals.')
if not isinstance(first_stage_anchor_generator,
grid_anchor_generator.GridAnchorGenerator):
raise ValueError('first_stage_anchor_generator must be of type '
'grid_anchor_generator.GridAnchorGenerator.')
self._is_training = is_training
self._image_resizer_fn = image_resizer_fn
self._feature_extractor = feature_extractor
self._first_stage_only = first_stage_only
# The first class is reserved as background.
unmatched_cls_target = tf.constant(
[1] + self._num_classes * [0], dtype=tf.float32)
self._proposal_target_assigner = target_assigner.create_target_assigner(
'FasterRCNN', 'proposal')
self._detector_target_assigner = target_assigner.create_target_assigner(
'FasterRCNN', 'detection', unmatched_cls_target=unmatched_cls_target)
# Both proposal and detector target assigners use the same box coder
self._box_coder = self._proposal_target_assigner.box_coder
# (First stage) Region proposal network parameters
self._first_stage_anchor_generator = first_stage_anchor_generator
self._first_stage_atrous_rate = first_stage_atrous_rate
self._first_stage_box_predictor_arg_scope = (
first_stage_box_predictor_arg_scope)
self._first_stage_box_predictor_kernel_size = (
first_stage_box_predictor_kernel_size)
self._first_stage_box_predictor_depth = first_stage_box_predictor_depth
self._first_stage_minibatch_size = first_stage_minibatch_size
self._first_stage_sampler = sampler.BalancedPositiveNegativeSampler(
positive_fraction=first_stage_positive_balance_fraction)
self._first_stage_box_predictor = box_predictor.ConvolutionalBoxPredictor(
self._is_training, num_classes=1,
conv_hyperparams=self._first_stage_box_predictor_arg_scope,
min_depth=0, max_depth=0, num_layers_before_predictor=0,
use_dropout=False, dropout_keep_prob=1.0, kernel_size=1,
box_code_size=self._box_coder.code_size)
self._first_stage_nms_score_threshold = first_stage_nms_score_threshold
self._first_stage_nms_iou_threshold = first_stage_nms_iou_threshold
self._first_stage_max_proposals = first_stage_max_proposals
self._first_stage_localization_loss = (
losses.WeightedSmoothL1LocalizationLoss(anchorwise_output=True))
self._first_stage_objectness_loss = (
losses.WeightedSoftmaxClassificationLoss(anchorwise_output=True))
self._first_stage_loc_loss_weight = first_stage_localization_loss_weight
self._first_stage_obj_loss_weight = first_stage_objectness_loss_weight
# Per-region cropping parameters
self._initial_crop_size = initial_crop_size
self._maxpool_kernel_size = maxpool_kernel_size
self._maxpool_stride = maxpool_stride
self._mask_rcnn_box_predictor = second_stage_mask_rcnn_box_predictor
self._second_stage_batch_size = second_stage_batch_size
self._second_stage_sampler = sampler.BalancedPositiveNegativeSampler(
positive_fraction=second_stage_balance_fraction)
self._second_stage_nms_fn = second_stage_non_max_suppression_fn
self._second_stage_score_conversion_fn = second_stage_score_conversion_fn
self._second_stage_localization_loss = (
losses.WeightedSmoothL1LocalizationLoss(anchorwise_output=True))
self._second_stage_classification_loss = (
losses.WeightedSoftmaxClassificationLoss(anchorwise_output=True))
self._second_stage_loc_loss_weight = second_stage_localization_loss_weight
self._second_stage_cls_loss_weight = second_stage_classification_loss_weight
self._hard_example_miner = hard_example_miner
self._parallel_iterations = parallel_iterations
@property
def first_stage_feature_extractor_scope(self):
return 'FirstStageFeatureExtractor'
@property
def second_stage_feature_extractor_scope(self):
return 'SecondStageFeatureExtractor'
@property
def first_stage_box_predictor_scope(self):
return 'FirstStageBoxPredictor'
@property
def second_stage_box_predictor_scope(self):
return 'SecondStageBoxPredictor'
@property
def max_num_proposals(self):
"""Max number of proposals (to pad to) for each image in the input batch.
At training time, this is set to be the `second_stage_batch_size` if hard
example miner is not configured, else it is set to
`first_stage_max_proposals`. At inference time, this is always set to
`first_stage_max_proposals`.
Returns:
A positive integer.
"""
if self._is_training and not self._hard_example_miner:
return self._second_stage_batch_size
return self._first_stage_max_proposals
def preprocess(self, inputs):
"""Feature-extractor specific preprocessing.
See base class.
For Faster R-CNN, we perform image resizing in the base class --- each
class subclassing FasterRCNNMetaArch is responsible for any additional
preprocessing (e.g., scaling pixel values to be in [-1, 1]).
Args:
inputs: a [batch, height_in, width_in, channels] float tensor representing
a batch of images with values between 0 and 255.0.
Returns:
preprocessed_inputs: a [batch, height_out, width_out, channels] float
tensor representing a batch of images.
Raises:
ValueError: if inputs tensor does not have type tf.float32
"""
if inputs.dtype is not tf.float32:
raise ValueError('`preprocess` expects a tf.float32 tensor')
with tf.name_scope('Preprocessor'):
resized_inputs = tf.map_fn(self._image_resizer_fn,
elems=inputs,
dtype=tf.float32,
parallel_iterations=self._parallel_iterations)
return self._feature_extractor.preprocess(resized_inputs)
def predict(self, preprocessed_inputs):
"""Predicts unpostprocessed tensors from input tensor.
This function takes an input batch of images and runs it through the
forward pass of the network to yield "raw" un-postprocessed predictions.
If `first_stage_only` is True, this function only returns first stage
RPN predictions (un-postprocessed). Otherwise it returns both
first stage RPN predictions as well as second stage box classifier
predictions.
Other remarks:
+ Anchor pruning vs. clipping: following the recommendation of the Faster
R-CNN paper, we prune anchors that venture outside the image window at
training time and clip anchors to the image window at inference time.
+ Proposal padding: as described at the top of the file, proposals are
padded to self._max_num_proposals and flattened so that proposals from all
images within the input batch are arranged along the same batch dimension.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
prediction_dict: a dictionary holding "raw" prediction tensors:
1) rpn_box_predictor_features: A 4-D float32 tensor with shape
[batch_size, height, width, depth] to be used for predicting proposal
boxes and corresponding objectness scores.
2) rpn_features_to_crop: A 4-D float32 tensor with shape
[batch_size, height, width, depth] representing image features to crop
using the proposal boxes predicted by the RPN.
3) image_shape: a 1-D tensor of shape [4] representing the input
image shape.
4) rpn_box_encodings: 3-D float tensor of shape
[batch_size, num_anchors, self._box_coder.code_size] containing
predicted boxes.
5) rpn_objectness_predictions_with_background: 3-D float tensor of shape
[batch_size, num_anchors, 2] containing class
predictions (logits) for each of the anchors. Note that this
tensor *includes* background class predictions (at class index 0).
6) anchors: A 2-D tensor of shape [num_anchors, 4] representing anchors
for the first stage RPN (in absolute coordinates). Note that
`num_anchors` can differ depending on whether the model is created in
training or inference mode.
(and if first_stage_only=False):
7) refined_box_encodings: a 3-D tensor with shape
[total_num_proposals, num_classes, 4] representing predicted
(final) refined box encodings, where
total_num_proposals=batch_size*self._max_num_proposals
8) class_predictions_with_background: a 3-D tensor with shape
[total_num_proposals, num_classes + 1] containing class
predictions (logits) for each of the anchors, where
total_num_proposals=batch_size*self._max_num_proposals.
Note that this tensor *includes* background class predictions
(at class index 0).
9) num_proposals: An int32 tensor of shape [batch_size] representing the
number of proposals generated by the RPN. `num_proposals` allows us
to keep track of which entries are to be treated as zero paddings and
which are not since we always pad the number of proposals to be
`self.max_num_proposals` for each image.
10) proposal_boxes: A float32 tensor of shape
[batch_size, self.max_num_proposals, 4] representing
decoded proposal bounding boxes (in absolute coordinates).
11) mask_predictions: (optional) a 4-D tensor with shape
[total_num_padded_proposals, num_classes, mask_height, mask_width]
containing instance mask predictions.
"""
(rpn_box_predictor_features, rpn_features_to_crop, anchors_boxlist,
image_shape) = self._extract_rpn_feature_maps(preprocessed_inputs)
(rpn_box_encodings, rpn_objectness_predictions_with_background
) = self._predict_rpn_proposals(rpn_box_predictor_features)
# The Faster R-CNN paper recommends pruning anchors that venture outside
# the image window at training time and clipping at inference time.
clip_window = tf.to_float(tf.stack([0, 0, image_shape[1], image_shape[2]]))
if self._is_training:
(rpn_box_encodings, rpn_objectness_predictions_with_background,
anchors_boxlist) = self._remove_invalid_anchors_and_predictions(
rpn_box_encodings, rpn_objectness_predictions_with_background,
anchors_boxlist, clip_window)
else:
anchors_boxlist = box_list_ops.clip_to_window(
anchors_boxlist, clip_window)
anchors = anchors_boxlist.get()
prediction_dict = {
'rpn_box_predictor_features': rpn_box_predictor_features,
'rpn_features_to_crop': rpn_features_to_crop,
'image_shape': image_shape,
'rpn_box_encodings': rpn_box_encodings,
'rpn_objectness_predictions_with_background':
rpn_objectness_predictions_with_background,
'anchors': anchors
}
if not self._first_stage_only:
prediction_dict.update(self._predict_second_stage(
rpn_box_encodings,
rpn_objectness_predictions_with_background,
rpn_features_to_crop,
anchors, image_shape))
return prediction_dict
def _predict_second_stage(self, rpn_box_encodings,
rpn_objectness_predictions_with_background,
rpn_features_to_crop,
anchors,
image_shape):
"""Predicts the output tensors from second stage of Faster R-CNN.
Args:
rpn_box_encodings: 4-D float tensor of shape
[batch_size, num_valid_anchors, self._box_coder.code_size] containing
predicted boxes.
rpn_objectness_predictions_with_background: 2-D float tensor of shape
[batch_size, num_valid_anchors, 2] containing class
predictions (logits) for each of the anchors. Note that this
tensor *includes* background class predictions (at class index 0).
rpn_features_to_crop: A 4-D float32 tensor with shape
[batch_size, height, width, depth] representing image features to crop
using the proposal boxes predicted by the RPN.
anchors: 2-D float tensor of shape
[num_anchors, self._box_coder.code_size].
image_shape: A 1D int32 tensors of size [4] containing the image shape.
Returns:
prediction_dict: a dictionary holding "raw" prediction tensors:
1) refined_box_encodings: a 3-D tensor with shape
[total_num_proposals, num_classes, 4] representing predicted
(final) refined box encodings, where
total_num_proposals=batch_size*self._max_num_proposals
2) class_predictions_with_background: a 3-D tensor with shape
[total_num_proposals, num_classes + 1] containing class
predictions (logits) for each of the anchors, where
total_num_proposals=batch_size*self._max_num_proposals.
Note that this tensor *includes* background class predictions
(at class index 0).
3) num_proposals: An int32 tensor of shape [batch_size] representing the
number of proposals generated by the RPN. `num_proposals` allows us
to keep track of which entries are to be treated as zero paddings and
which are not since we always pad the number of proposals to be
`self.max_num_proposals` for each image.
4) proposal_boxes: A float32 tensor of shape
[batch_size, self.max_num_proposals, 4] representing
decoded proposal bounding boxes (in absolute coordinates).
5) mask_predictions: (optional) a 4-D tensor with shape
[total_num_padded_proposals, num_classes, mask_height, mask_width]
containing instance mask predictions.
"""
proposal_boxes_normalized, _, num_proposals = self._postprocess_rpn(
rpn_box_encodings, rpn_objectness_predictions_with_background,
anchors, image_shape)
flattened_proposal_feature_maps = (
self._compute_second_stage_input_feature_maps(
rpn_features_to_crop, proposal_boxes_normalized))
box_classifier_features = (
self._feature_extractor.extract_box_classifier_features(
flattened_proposal_feature_maps,
scope=self.second_stage_feature_extractor_scope))
box_predictions = self._mask_rcnn_box_predictor.predict(
box_classifier_features,
num_predictions_per_location=1,
scope=self.second_stage_box_predictor_scope)
refined_box_encodings = tf.squeeze(
box_predictions[box_predictor.BOX_ENCODINGS], axis=1)
class_predictions_with_background = tf.squeeze(box_predictions[
box_predictor.CLASS_PREDICTIONS_WITH_BACKGROUND], axis=1)
absolute_proposal_boxes = ops.normalized_to_image_coordinates(
proposal_boxes_normalized, image_shape, self._parallel_iterations)
prediction_dict = {
'refined_box_encodings': refined_box_encodings,
'class_predictions_with_background':
class_predictions_with_background,
'num_proposals': num_proposals,
'proposal_boxes': absolute_proposal_boxes,
}
return prediction_dict
def _extract_rpn_feature_maps(self, preprocessed_inputs):
"""Extracts RPN features.
This function extracts two feature maps: a feature map to be directly
fed to a box predictor (to predict location and objectness scores for
proposals) and a feature map from which to crop regions which will then
be sent to the second stage box classifier.
Args:
preprocessed_inputs: a [batch, height, width, channels] image tensor.
Returns:
rpn_box_predictor_features: A 4-D float32 tensor with shape
[batch, height, width, depth] to be used for predicting proposal boxes
and corresponding objectness scores.
rpn_features_to_crop: A 4-D float32 tensor with shape
[batch, height, width, depth] representing image features to crop using
the proposals boxes.
anchors: A BoxList representing anchors (for the RPN) in
absolute coordinates.
image_shape: A 1-D tensor representing the input image shape.
"""
image_shape = tf.shape(preprocessed_inputs)
rpn_features_to_crop = self._feature_extractor.extract_proposal_features(
preprocessed_inputs, scope=self.first_stage_feature_extractor_scope)
feature_map_shape = tf.shape(rpn_features_to_crop)
anchors = self._first_stage_anchor_generator.generate(
[(feature_map_shape[1], feature_map_shape[2])])
with slim.arg_scope(self._first_stage_box_predictor_arg_scope):
kernel_size = self._first_stage_box_predictor_kernel_size
rpn_box_predictor_features = slim.conv2d(
rpn_features_to_crop,
self._first_stage_box_predictor_depth,
kernel_size=[kernel_size, kernel_size],
rate=self._first_stage_atrous_rate,
activation_fn=tf.nn.relu6)
return (rpn_box_predictor_features, rpn_features_to_crop,
anchors, image_shape)
def _predict_rpn_proposals(self, rpn_box_predictor_features):
"""Adds box predictors to RPN feature map to predict proposals.
Note resulting tensors will not have been postprocessed.
Args:
rpn_box_predictor_features: A 4-D float32 tensor with shape
[batch, height, width, depth] to be used for predicting proposal boxes
and corresponding objectness scores.
Returns:
box_encodings: 3-D float tensor of shape
[batch_size, num_anchors, self._box_coder.code_size] containing
predicted boxes.
objectness_predictions_with_background: 3-D float tensor of shape
[batch_size, num_anchors, 2] containing class
predictions (logits) for each of the anchors. Note that this
tensor *includes* background class predictions (at class index 0).
Raises:
RuntimeError: if the anchor generator generates anchors corresponding to
multiple feature maps. We currently assume that a single feature map
is generated for the RPN.
"""
num_anchors_per_location = (
self._first_stage_anchor_generator.num_anchors_per_location())
if len(num_anchors_per_location) != 1:
raise RuntimeError('anchor_generator is expected to generate anchors '
'corresponding to a single feature map.')
box_predictions = self._first_stage_box_predictor.predict(
rpn_box_predictor_features,
num_anchors_per_location[0],
scope=self.first_stage_box_predictor_scope)
box_encodings = box_predictions[box_predictor.BOX_ENCODINGS]
objectness_predictions_with_background = box_predictions[
box_predictor.CLASS_PREDICTIONS_WITH_BACKGROUND]
return (tf.squeeze(box_encodings, axis=2),
objectness_predictions_with_background)
def _remove_invalid_anchors_and_predictions(
self,
box_encodings,
objectness_predictions_with_background,
anchors_boxlist,
clip_window):
"""Removes anchors that (partially) fall outside an image.
Also removes associated box encodings and objectness predictions.
Args:
box_encodings: 3-D float tensor of shape
[batch_size, num_anchors, self._box_coder.code_size] containing
predicted boxes.
objectness_predictions_with_background: 3-D float tensor of shape
[batch_size, num_anchors, 2] containing class
predictions (logits) for each of the anchors. Note that this
tensor *includes* background class predictions (at class index 0).
anchors_boxlist: A BoxList representing num_anchors anchors (for the RPN)
in absolute coordinates.
clip_window: a 1-D tensor representing the [ymin, xmin, ymax, xmax]
extent of the window to clip/prune to.
Returns:
box_encodings: 4-D float tensor of shape
[batch_size, num_valid_anchors, self._box_coder.code_size] containing
predicted boxes, where num_valid_anchors <= num_anchors
objectness_predictions_with_background: 2-D float tensor of shape
[batch_size, num_valid_anchors, 2] containing class
predictions (logits) for each of the anchors, where
num_valid_anchors <= num_anchors. Note that this
tensor *includes* background class predictions (at class index 0).
anchors: A BoxList representing num_valid_anchors anchors (for the RPN) in
absolute coordinates.
"""
pruned_anchors_boxlist, keep_indices = box_list_ops.prune_outside_window(
anchors_boxlist, clip_window)
def _batch_gather_kept_indices(predictions_tensor):
return tf.map_fn(
partial(tf.gather, indices=keep_indices),
elems=predictions_tensor,
dtype=tf.float32,
parallel_iterations=self._parallel_iterations,
back_prop=True)
return (_batch_gather_kept_indices(box_encodings),
_batch_gather_kept_indices(objectness_predictions_with_background),
pruned_anchors_boxlist)
def _flatten_first_two_dimensions(self, inputs):
"""Flattens `K-d` tensor along batch dimension to be a `(K-1)-d` tensor.
Converts `inputs` with shape [A, B, ..., depth] into a tensor of shape
[A * B, ..., depth].
Args:
inputs: A float tensor with shape [A, B, ..., depth]. Note that the first
two and last dimensions must be statically defined.
Returns:
A float tensor with shape [A * B, ..., depth] (where the first and last
dimension are statically defined.
"""
combined_shape = shape_utils.combined_static_and_dynamic_shape(inputs)
flattened_shape = tf.stack([combined_shape[0] * combined_shape[1]] +
combined_shape[2:])
return tf.reshape(inputs, flattened_shape)
def postprocess(self, prediction_dict):
"""Convert prediction tensors to final detections.
This function converts raw predictions tensors to final detection results.
See base class for output format conventions. Note also that by default,
scores are to be interpreted as logits, but if a score_converter is used,
then scores are remapped (and may thus have a different interpretation).
If first_stage_only=True, the returned results represent proposals from the
first stage RPN and are padded to have self.max_num_proposals for each
image; otherwise, the results can be interpreted as multiclass detections
from the full two-stage model and are padded to self._max_detections.
Args:
prediction_dict: a dictionary holding prediction tensors (see the
documentation for the predict method. If first_stage_only=True, we
expect prediction_dict to contain `rpn_box_encodings`,
`rpn_objectness_predictions_with_background`, `rpn_features_to_crop`,
`image_shape`, and `anchors` fields. Otherwise we expect
prediction_dict to additionally contain `refined_box_encodings`,
`class_predictions_with_background`, `num_proposals`,
`proposal_boxes` and, optionally, `mask_predictions` fields.
Returns:
detections: a dictionary containing the following fields
detection_boxes: [batch, max_detection, 4]
detection_scores: [batch, max_detections]
detection_classes: [batch, max_detections]
(this entry is only created if rpn_mode=False)
num_detections: [batch]
"""
with tf.name_scope('FirstStagePostprocessor'):
image_shape = prediction_dict['image_shape']
if self._first_stage_only:
proposal_boxes, proposal_scores, num_proposals = self._postprocess_rpn(
prediction_dict['rpn_box_encodings'],
prediction_dict['rpn_objectness_predictions_with_background'],
prediction_dict['anchors'],
image_shape)
return {
'detection_boxes': proposal_boxes,
'detection_scores': proposal_scores,
'num_detections': num_proposals
}
with tf.name_scope('SecondStagePostprocessor'):
mask_predictions = prediction_dict.get(box_predictor.MASK_PREDICTIONS)
detections_dict = self._postprocess_box_classifier(
prediction_dict['refined_box_encodings'],
prediction_dict['class_predictions_with_background'],
prediction_dict['proposal_boxes'],
prediction_dict['num_proposals'],
image_shape,
mask_predictions=mask_predictions)
return detections_dict
def _postprocess_rpn(self,
rpn_box_encodings_batch,
rpn_objectness_predictions_with_background_batch,
anchors,
image_shape):
"""Converts first stage prediction tensors from the RPN to proposals.
This function decodes the raw RPN predictions, runs non-max suppression
on the result.
Note that the behavior of this function is slightly modified during
training --- specifically, we stop the gradient from passing through the
proposal boxes and we only return a balanced sampled subset of proposals
with size `second_stage_batch_size`.
Args:
rpn_box_encodings_batch: A 3-D float32 tensor of shape
[batch_size, num_anchors, self._box_coder.code_size] containing
predicted proposal box encodings.
rpn_objectness_predictions_with_background_batch: A 3-D float tensor of
shape [batch_size, num_anchors, 2] containing objectness predictions
(logits) for each of the anchors with 0 corresponding to background
and 1 corresponding to object.
anchors: A 2-D tensor of shape [num_anchors, 4] representing anchors
for the first stage RPN. Note that `num_anchors` can differ depending
on whether the model is created in training or inference mode.
image_shape: A 1-D tensor representing the input image shape.
Returns:
proposal_boxes: A float tensor with shape
[batch_size, max_num_proposals, 4] representing the (potentially zero
padded) proposal boxes for all images in the batch. These boxes are
represented as normalized coordinates.
proposal_scores: A float tensor with shape
[batch_size, max_num_proposals] representing the (potentially zero
padded) proposal objectness scores for all images in the batch.
num_proposals: A Tensor of type `int32`. A 1-D tensor of shape [batch]
representing the number of proposals predicted for each image in
the batch.
"""
rpn_box_encodings_batch = tf.expand_dims(rpn_box_encodings_batch, axis=2)
rpn_encodings_shape = shape_utils.combined_static_and_dynamic_shape(
rpn_box_encodings_batch)
tiled_anchor_boxes = tf.tile(
tf.expand_dims(anchors, 0), [rpn_encodings_shape[0], 1, 1])
proposal_boxes = self._batch_decode_boxes(rpn_box_encodings_batch,
tiled_anchor_boxes)
proposal_boxes = tf.squeeze(proposal_boxes, axis=2)
rpn_objectness_softmax_without_background = tf.nn.softmax(
rpn_objectness_predictions_with_background_batch)[:, :, 1]
clip_window = tf.to_float(tf.stack([0, 0, image_shape[1], image_shape[2]]))
(proposal_boxes, proposal_scores, _, _,
num_proposals) = post_processing.batch_multiclass_non_max_suppression(
tf.expand_dims(proposal_boxes, axis=2),
tf.expand_dims(rpn_objectness_softmax_without_background,
axis=2),
self._first_stage_nms_score_threshold,
self._first_stage_nms_iou_threshold,
self._first_stage_max_proposals,
self._first_stage_max_proposals,
clip_window=clip_window)
if self._is_training:
proposal_boxes = tf.stop_gradient(proposal_boxes)
if not self._hard_example_miner:
(groundtruth_boxlists, groundtruth_classes_with_background_list,
) = self._format_groundtruth_data(image_shape)
(proposal_boxes, proposal_scores,
num_proposals) = self._unpad_proposals_and_sample_box_classifier_batch(
proposal_boxes, proposal_scores, num_proposals,
groundtruth_boxlists, groundtruth_classes_with_background_list)
# normalize proposal boxes
proposal_boxes_reshaped = tf.reshape(proposal_boxes, [-1, 4])
normalized_proposal_boxes_reshaped = box_list_ops.to_normalized_coordinates(
box_list.BoxList(proposal_boxes_reshaped),
image_shape[1], image_shape[2], check_range=False).get()
proposal_boxes = tf.reshape(normalized_proposal_boxes_reshaped,
[-1, proposal_boxes.shape[1].value, 4])
return proposal_boxes, proposal_scores, num_proposals
def _unpad_proposals_and_sample_box_classifier_batch(
self,
proposal_boxes,
proposal_scores,
num_proposals,
groundtruth_boxlists,
groundtruth_classes_with_background_list):
"""Unpads proposals and samples a minibatch for second stage.
Args:
proposal_boxes: A float tensor with shape
[batch_size, num_proposals, 4] representing the (potentially zero
padded) proposal boxes for all images in the batch. These boxes are
represented as normalized coordinates.
proposal_scores: A float tensor with shape
[batch_size, num_proposals] representing the (potentially zero
padded) proposal objectness scores for all images in the batch.
num_proposals: A Tensor of type `int32`. A 1-D tensor of shape [batch]
representing the number of proposals predicted for each image in
the batch.
groundtruth_boxlists: A list of BoxLists containing (absolute) coordinates
of the groundtruth boxes.
groundtruth_classes_with_background_list: A list of 2-D one-hot
(or k-hot) tensors of shape [num_boxes, num_classes+1] containing the
class targets with the 0th index assumed to map to the background class.
Returns:
proposal_boxes: A float tensor with shape
[batch_size, second_stage_batch_size, 4] representing the (potentially
zero padded) proposal boxes for all images in the batch. These boxes
are represented as normalized coordinates.
proposal_scores: A float tensor with shape
[batch_size, second_stage_batch_size] representing the (potentially zero
padded) proposal objectness scores for all images in the batch.
num_proposals: A Tensor of type `int32`. A 1-D tensor of shape [batch]
representing the number of proposals predicted for each image in
the batch.
"""
single_image_proposal_box_sample = []
single_image_proposal_score_sample = []
single_image_num_proposals_sample = []
for (single_image_proposal_boxes,
single_image_proposal_scores,
single_image_num_proposals,
single_image_groundtruth_boxlist,
single_image_groundtruth_classes_with_background) in zip(
tf.unstack(proposal_boxes),
tf.unstack(proposal_scores),
tf.unstack(num_proposals),
groundtruth_boxlists,
groundtruth_classes_with_background_list):
static_shape = single_image_proposal_boxes.get_shape()
sliced_static_shape = tf.TensorShape([tf.Dimension(None),
static_shape.dims[-1]])
single_image_proposal_boxes = tf.slice(
single_image_proposal_boxes,
[0, 0],
[single_image_num_proposals, -1])
single_image_proposal_boxes.set_shape(sliced_static_shape)
single_image_proposal_scores = tf.slice(single_image_proposal_scores,
[0],
[single_image_num_proposals])
single_image_boxlist = box_list.BoxList(single_image_proposal_boxes)
single_image_boxlist.add_field(fields.BoxListFields.scores,
single_image_proposal_scores)
sampled_boxlist = self._sample_box_classifier_minibatch(
single_image_boxlist,
single_image_groundtruth_boxlist,
single_image_groundtruth_classes_with_background)
sampled_padded_boxlist = box_list_ops.pad_or_clip_box_list(
sampled_boxlist,
num_boxes=self._second_stage_batch_size)
single_image_num_proposals_sample.append(tf.minimum(
sampled_boxlist.num_boxes(),
self._second_stage_batch_size))
bb = sampled_padded_boxlist.get()
single_image_proposal_box_sample.append(bb)
single_image_proposal_score_sample.append(
sampled_padded_boxlist.get_field(fields.BoxListFields.scores))
return (tf.stack(single_image_proposal_box_sample),
tf.stack(single_image_proposal_score_sample),
tf.stack(single_image_num_proposals_sample))
def _format_groundtruth_data(self, image_shape):
"""Helper function for preparing groundtruth data for target assignment.
In order to be consistent with the model.DetectionModel interface,
groundtruth boxes are specified in normalized coordinates and classes are
specified as label indices with no assumed background category. To prepare
for target assignment, we:
1) convert boxes to absolute coordinates,
2) add a background class at class index 0