coco_evaluation.py 31.1 KB
Newer Older
Zhichao Lu's avatar
Zhichao Lu committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# 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.
# ==============================================================================
15
16
17
18
19
20
21
22
23
24
25
26
"""Class for evaluating object detections with COCO metrics."""
import numpy as np
import tensorflow as tf

from object_detection.core import standard_fields
from object_detection.metrics import coco_tools
from object_detection.utils import object_detection_evaluation


class CocoDetectionEvaluator(object_detection_evaluation.DetectionEvaluator):
  """Class to evaluate COCO detection metrics."""

27
28
29
30
  def __init__(self,
               categories,
               include_metrics_per_category=False,
               all_metrics_per_category=False):
31
32
33
34
35
36
    """Constructor.

    Args:
      categories: A list of dicts, each of which has the following keys -
        'id': (required) an integer id uniquely identifying this category.
        'name': (required) string representing category name e.g., 'cat', 'dog'.
37
      include_metrics_per_category: If True, include metrics for each category.
38
39
40
41
42
43
44
45
46
47
48
49
50
51
      all_metrics_per_category: Whether to include all the summary metrics for
        each category in per_category_ap. Be careful with setting it to true if
        you have more than handful of categories, because it will pollute
        your mldash.
    """
    super(CocoDetectionEvaluator, self).__init__(categories)
    # _image_ids is a dictionary that maps unique image ids to Booleans which
    # indicate whether a corresponding detection has been added.
    self._image_ids = {}
    self._groundtruth_list = []
    self._detection_boxes_list = []
    self._category_id_set = set([cat['id'] for cat in self._categories])
    self._annotation_id = 1
    self._metrics = None
52
    self._include_metrics_per_category = include_metrics_per_category
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
    self._all_metrics_per_category = all_metrics_per_category

  def clear(self):
    """Clears the state to prepare for a fresh evaluation."""
    self._image_ids.clear()
    self._groundtruth_list = []
    self._detection_boxes_list = []

  def add_single_ground_truth_image_info(self,
                                         image_id,
                                         groundtruth_dict):
    """Adds groundtruth for a single image to be used for evaluation.

    If the image has already been added, a warning is logged, and groundtruth is
    ignored.

    Args:
      image_id: A unique string/integer identifier for the image.
      groundtruth_dict: A dictionary containing -
        InputDataFields.groundtruth_boxes: float32 numpy array of shape
          [num_boxes, 4] containing `num_boxes` groundtruth boxes of the format
          [ymin, xmin, ymax, xmax] in absolute image coordinates.
        InputDataFields.groundtruth_classes: integer numpy array of shape
          [num_boxes] containing 1-indexed groundtruth classes for the boxes.
77
78
        InputDataFields.groundtruth_is_crowd (optional): integer numpy array of
          shape [num_boxes] containing iscrowd flag for groundtruth boxes.
79
80
81
82
83
84
    """
    if image_id in self._image_ids:
      tf.logging.warning('Ignoring ground truth with image id %s since it was '
                         'previously added', image_id)
      return

85
86
87
88
89
90
    groundtruth_is_crowd = groundtruth_dict.get(
        standard_fields.InputDataFields.groundtruth_is_crowd)
    # Drop groundtruth_is_crowd if empty tensor.
    if groundtruth_is_crowd is not None and not groundtruth_is_crowd.shape[0]:
      groundtruth_is_crowd = None

91
    self._groundtruth_list.extend(
92
        coco_tools.ExportSingleImageGroundtruthToCoco(
93
94
95
            image_id=image_id,
            next_annotation_id=self._annotation_id,
            category_id_set=self._category_id_set,
96
97
98
99
100
            groundtruth_boxes=groundtruth_dict[
                standard_fields.InputDataFields.groundtruth_boxes],
            groundtruth_classes=groundtruth_dict[
                standard_fields.InputDataFields.groundtruth_classes],
            groundtruth_is_crowd=groundtruth_is_crowd))
101
102
    self._annotation_id += groundtruth_dict[standard_fields.InputDataFields.
                                            groundtruth_boxes].shape[0]
103
    # Boolean to indicate whether a detection has been added for this image.
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
    self._image_ids[image_id] = False

  def add_single_detected_image_info(self,
                                     image_id,
                                     detections_dict):
    """Adds detections for a single image to be used for evaluation.

    If a detection has already been added for this image id, a warning is
    logged, and the detection is skipped.

    Args:
      image_id: A unique string/integer identifier for the image.
      detections_dict: A dictionary containing -
        DetectionResultFields.detection_boxes: float32 numpy array of shape
          [num_boxes, 4] containing `num_boxes` detection boxes of the format
          [ymin, xmin, ymax, xmax] in absolute image coordinates.
        DetectionResultFields.detection_scores: float32 numpy array of shape
          [num_boxes] containing detection scores for the boxes.
        DetectionResultFields.detection_classes: integer numpy array of shape
          [num_boxes] containing 1-indexed detection classes for the boxes.

    Raises:
      ValueError: If groundtruth for the image_id is not available.
    """
    if image_id not in self._image_ids:
      raise ValueError('Missing groundtruth for image id: {}'.format(image_id))

    if self._image_ids[image_id]:
      tf.logging.warning('Ignoring detection with image id %s since it was '
                         'previously added', image_id)
      return

    self._detection_boxes_list.extend(
        coco_tools.ExportSingleImageDetectionBoxesToCoco(
            image_id=image_id,
            category_id_set=self._category_id_set,
            detection_boxes=detections_dict[standard_fields.
                                            DetectionResultFields
                                            .detection_boxes],
            detection_scores=detections_dict[standard_fields.
                                             DetectionResultFields.
                                             detection_scores],
            detection_classes=detections_dict[standard_fields.
                                              DetectionResultFields.
                                              detection_classes]))
    self._image_ids[image_id] = True

  def evaluate(self):
    """Evaluates the detection boxes and returns a dictionary of coco metrics.

    Returns:
      A dictionary holding -

      1. summary_metrics:
      'DetectionBoxes_Precision/mAP': mean average precision over classes
        averaged over IOU thresholds ranging from .5 to .95 with .05
        increments.
      'DetectionBoxes_Precision/mAP@.50IOU': mean average precision at 50% IOU
      'DetectionBoxes_Precision/mAP@.75IOU': mean average precision at 75% IOU
      'DetectionBoxes_Precision/mAP (small)': mean average precision for small
        objects (area < 32^2 pixels).
      'DetectionBoxes_Precision/mAP (medium)': mean average precision for
        medium sized objects (32^2 pixels < area < 96^2 pixels).
      'DetectionBoxes_Precision/mAP (large)': mean average precision for large
        objects (96^2 pixels < area < 10000^2 pixels).
      'DetectionBoxes_Recall/AR@1': average recall with 1 detection.
      'DetectionBoxes_Recall/AR@10': average recall with 10 detections.
      'DetectionBoxes_Recall/AR@100': average recall with 100 detections.
      'DetectionBoxes_Recall/AR@100 (small)': average recall for small objects
        with 100.
      'DetectionBoxes_Recall/AR@100 (medium)': average recall for medium objects
        with 100.
      'DetectionBoxes_Recall/AR@100 (large)': average recall for large objects
        with 100 detections.

179
180
      2. per_category_ap: if include_metrics_per_category is True, category
      specific results with keys of the form:
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
      'Precision mAP ByCategory/category' (without the supercategory part if
      no supercategories exist). For backward compatibility
      'PerformanceByCategory' is included in the output regardless of
      all_metrics_per_category.
    """
    groundtruth_dict = {
        'annotations': self._groundtruth_list,
        'images': [{'id': image_id} for image_id in self._image_ids],
        'categories': self._categories
    }
    coco_wrapped_groundtruth = coco_tools.COCOWrapper(groundtruth_dict)
    coco_wrapped_detections = coco_wrapped_groundtruth.LoadAnnotations(
        self._detection_boxes_list)
    box_evaluator = coco_tools.COCOEvalWrapper(
        coco_wrapped_groundtruth, coco_wrapped_detections, agnostic_mode=False)
    box_metrics, box_per_category_ap = box_evaluator.ComputeMetrics(
197
        include_metrics_per_category=self._include_metrics_per_category,
198
199
200
        all_metrics_per_category=self._all_metrics_per_category)
    box_metrics.update(box_per_category_ap)
    box_metrics = {'DetectionBoxes_'+ key: value
201
                   for key, value in iter(box_metrics.items())}
202
203
204
    return box_metrics

  def get_estimator_eval_metric_ops(self, image_id, groundtruth_boxes,
205
206
                                    groundtruth_classes,
                                    detection_boxes,
207
                                    detection_scores, detection_classes,
208
                                    groundtruth_is_crowd=None,
209
210
                                    num_gt_boxes_per_image=None,
                                    num_det_boxes_per_image=None):
211
212
213
214
215
    """Returns a dictionary of eval metric ops to use with `tf.EstimatorSpec`.

    Note that once value_op is called, the detections and groundtruth added via
    update_op are cleared.

216
217
218
219
    This function can take in groundtruth and detections for a batch of images,
    or for a single image. For the latter case, the batch dimension for input
    tensors need not be present.

220
    Args:
221
222
223
224
      image_id: string/integer tensor of shape [batch] with unique identifiers
        for the images.
      groundtruth_boxes: float32 tensor of shape [batch, num_boxes, 4]
        containing `num_boxes` groundtruth boxes of the format
225
        [ymin, xmin, ymax, xmax] in absolute image coordinates.
226
      groundtruth_classes: int32 tensor of shape [batch, num_boxes] containing
227
        1-indexed groundtruth classes for the boxes.
228
      detection_boxes: float32 tensor of shape [batch, num_boxes, 4] containing
229
230
        `num_boxes` detection boxes of the format [ymin, xmin, ymax, xmax]
        in absolute image coordinates.
231
      detection_scores: float32 tensor of shape [batch, num_boxes] containing
232
        detection scores for the boxes.
233
      detection_classes: int32 tensor of shape [batch, num_boxes] containing
234
        1-indexed detection classes for the boxes.
235
236
237
      groundtruth_is_crowd: bool tensor of shape [batch, num_boxes] containing
        is_crowd annotations. This field is optional, and if not passed, then
        all boxes are treated as *not* is_crowd.
238
239
240
241
242
243
      num_gt_boxes_per_image: int32 tensor of shape [batch] containing the
        number of groundtruth boxes per image. If None, will assume no padding
        in groundtruth tensors.
      num_det_boxes_per_image: int32 tensor of shape [batch] containing the
        number of detection boxes per image. If None, will assume no padding in
        the detection tensors.
244
245
246
247
248
249
250
251

    Returns:
      a dictionary of metric names to tuple of value_op and update_op that can
      be used as eval metric ops in tf.EstimatorSpec. Note that all update ops
      must be run together and similarly all value ops must be run together to
      guarantee correct behaviour.
    """
    def update_op(
252
253
254
        image_id_batched,
        groundtruth_boxes_batched,
        groundtruth_classes_batched,
255
        groundtruth_is_crowd_batched,
256
257
258
259
260
261
262
        num_gt_boxes_per_image,
        detection_boxes_batched,
        detection_scores_batched,
        detection_classes_batched,
        num_det_boxes_per_image):
      """Update operation for adding batch of images to Coco evaluator."""

263
264
      for (image_id, gt_box, gt_class, gt_is_crowd, num_gt_box, det_box,
           det_score, det_class, num_det_box) in zip(
265
               image_id_batched, groundtruth_boxes_batched,
266
267
               groundtruth_classes_batched, groundtruth_is_crowd_batched,
               num_gt_boxes_per_image,
268
269
270
271
272
               detection_boxes_batched, detection_scores_batched,
               detection_classes_batched, num_det_boxes_per_image):
        self.add_single_ground_truth_image_info(
            image_id,
            {'groundtruth_boxes': gt_box[:num_gt_box],
273
274
             'groundtruth_classes': gt_class[:num_gt_box],
             'groundtruth_is_crowd': gt_is_crowd[:num_gt_box]})
275
276
277
278
279
280
        self.add_single_detected_image_info(
            image_id,
            {'detection_boxes': det_box[:num_det_box],
             'detection_scores': det_score[:num_det_box],
             'detection_classes': det_class[:num_det_box]})

281
282
    if groundtruth_is_crowd is None:
      groundtruth_is_crowd = tf.zeros_like(groundtruth_classes, dtype=tf.bool)
283
284
285
286
287
    if not image_id.shape.as_list():
      # Apply a batch dimension to all tensors.
      image_id = tf.expand_dims(image_id, 0)
      groundtruth_boxes = tf.expand_dims(groundtruth_boxes, 0)
      groundtruth_classes = tf.expand_dims(groundtruth_classes, 0)
288
      groundtruth_is_crowd = tf.expand_dims(groundtruth_is_crowd, 0)
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
      detection_boxes = tf.expand_dims(detection_boxes, 0)
      detection_scores = tf.expand_dims(detection_scores, 0)
      detection_classes = tf.expand_dims(detection_classes, 0)

      if num_gt_boxes_per_image is None:
        num_gt_boxes_per_image = tf.shape(groundtruth_boxes)[1:2]
      else:
        num_gt_boxes_per_image = tf.expand_dims(num_gt_boxes_per_image, 0)

      if num_det_boxes_per_image is None:
        num_det_boxes_per_image = tf.shape(detection_boxes)[1:2]
      else:
        num_det_boxes_per_image = tf.expand_dims(num_det_boxes_per_image, 0)
    else:
      if num_gt_boxes_per_image is None:
        num_gt_boxes_per_image = tf.tile(
            tf.shape(groundtruth_boxes)[1:2],
            multiples=tf.shape(groundtruth_boxes)[0:1])
      if num_det_boxes_per_image is None:
        num_det_boxes_per_image = tf.tile(
            tf.shape(detection_boxes)[1:2],
            multiples=tf.shape(detection_boxes)[0:1])
311
312
313
314

    update_op = tf.py_func(update_op, [image_id,
                                       groundtruth_boxes,
                                       groundtruth_classes,
315
                                       groundtruth_is_crowd,
316
                                       num_gt_boxes_per_image,
317
318
                                       detection_boxes,
                                       detection_scores,
319
320
                                       detection_classes,
                                       num_det_boxes_per_image], [])
321
322
323
324
325
326
327
328
329
330
331
332
    metric_names = ['DetectionBoxes_Precision/mAP',
                    'DetectionBoxes_Precision/mAP@.50IOU',
                    'DetectionBoxes_Precision/mAP@.75IOU',
                    'DetectionBoxes_Precision/mAP (large)',
                    'DetectionBoxes_Precision/mAP (medium)',
                    'DetectionBoxes_Precision/mAP (small)',
                    'DetectionBoxes_Recall/AR@1',
                    'DetectionBoxes_Recall/AR@10',
                    'DetectionBoxes_Recall/AR@100',
                    'DetectionBoxes_Recall/AR@100 (large)',
                    'DetectionBoxes_Recall/AR@100 (medium)',
                    'DetectionBoxes_Recall/AR@100 (small)']
333
334
335
336
    if self._include_metrics_per_category:
      for category_dict in self._categories:
        metric_names.append('DetectionBoxes_PerformanceByCategory/mAP/' +
                            category_dict['name'])
337
338
339
340
341
342
343
344
345
346
347

    def first_value_func():
      self._metrics = self.evaluate()
      self.clear()
      return np.float32(self._metrics[metric_names[0]])

    def value_func_factory(metric_name):
      def value_func():
        return np.float32(self._metrics[metric_name])
      return value_func

348
    # Ensure that the metrics are only evaluated once.
349
350
351
352
353
354
355
356
357
358
    first_value_op = tf.py_func(first_value_func, [], tf.float32)
    eval_metric_ops = {metric_names[0]: (first_value_op, update_op)}
    with tf.control_dependencies([first_value_op]):
      for metric_name in metric_names[1:]:
        eval_metric_ops[metric_name] = (tf.py_func(
            value_func_factory(metric_name), [], np.float32), update_op)
    return eval_metric_ops


def _check_mask_type_and_value(array_name, masks):
359
  """Checks whether mask dtype is uint8 and the values are either 0 or 1."""
360
361
362
363
364
365
366
367
368
369
370
  if masks.dtype != np.uint8:
    raise ValueError('{} must be of type np.uint8. Found {}.'.format(
        array_name, masks.dtype))
  if np.any(np.logical_and(masks != 0, masks != 1)):
    raise ValueError('{} elements can only be either 0 or 1.'.format(
        array_name))


class CocoMaskEvaluator(object_detection_evaluation.DetectionEvaluator):
  """Class to evaluate COCO detection metrics."""

371
  def __init__(self, categories, include_metrics_per_category=False):
372
373
374
375
376
377
    """Constructor.

    Args:
      categories: A list of dicts, each of which has the following keys -
        'id': (required) an integer id uniquely identifying this category.
        'name': (required) string representing category name e.g., 'cat', 'dog'.
378
      include_metrics_per_category: If True, include metrics for each category.
379
380
381
382
383
384
385
386
    """
    super(CocoMaskEvaluator, self).__init__(categories)
    self._image_id_to_mask_shape_map = {}
    self._image_ids_with_detections = set([])
    self._groundtruth_list = []
    self._detection_masks_list = []
    self._category_id_set = set([cat['id'] for cat in self._categories])
    self._annotation_id = 1
387
    self._include_metrics_per_category = include_metrics_per_category
388
389
390
391
392
393
394
395
396
397
398
399
400

  def clear(self):
    """Clears the state to prepare for a fresh evaluation."""
    self._image_id_to_mask_shape_map.clear()
    self._image_ids_with_detections.clear()
    self._groundtruth_list = []
    self._detection_masks_list = []

  def add_single_ground_truth_image_info(self,
                                         image_id,
                                         groundtruth_dict):
    """Adds groundtruth for a single image to be used for evaluation.

401
402
403
    If the image has already been added, a warning is logged, and groundtruth is
    ignored.

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
    Args:
      image_id: A unique string/integer identifier for the image.
      groundtruth_dict: A dictionary containing -
        InputDataFields.groundtruth_boxes: float32 numpy array of shape
          [num_boxes, 4] containing `num_boxes` groundtruth boxes of the format
          [ymin, xmin, ymax, xmax] in absolute image coordinates.
        InputDataFields.groundtruth_classes: integer numpy array of shape
          [num_boxes] containing 1-indexed groundtruth classes for the boxes.
        InputDataFields.groundtruth_instance_masks: uint8 numpy array of shape
          [num_boxes, image_height, image_width] containing groundtruth masks
          corresponding to the boxes. The elements of the array must be in
          {0, 1}.
    """
    if image_id in self._image_id_to_mask_shape_map:
      tf.logging.warning('Ignoring ground truth with image id %s since it was '
                         'previously added', image_id)
      return

    groundtruth_instance_masks = groundtruth_dict[
        standard_fields.InputDataFields.groundtruth_instance_masks]
    _check_mask_type_and_value(standard_fields.InputDataFields.
                               groundtruth_instance_masks,
                               groundtruth_instance_masks)
    self._groundtruth_list.extend(
        coco_tools.
        ExportSingleImageGroundtruthToCoco(
            image_id=image_id,
            next_annotation_id=self._annotation_id,
            category_id_set=self._category_id_set,
            groundtruth_boxes=groundtruth_dict[standard_fields.InputDataFields.
                                               groundtruth_boxes],
            groundtruth_classes=groundtruth_dict[standard_fields.
                                                 InputDataFields.
                                                 groundtruth_classes],
            groundtruth_masks=groundtruth_instance_masks))
    self._annotation_id += groundtruth_dict[standard_fields.InputDataFields.
                                            groundtruth_boxes].shape[0]
    self._image_id_to_mask_shape_map[image_id] = groundtruth_dict[
        standard_fields.InputDataFields.groundtruth_instance_masks].shape

  def add_single_detected_image_info(self,
                                     image_id,
                                     detections_dict):
    """Adds detections for a single image to be used for evaluation.

449
450
451
    If a detection has already been added for this image id, a warning is
    logged, and the detection is skipped.

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
    Args:
      image_id: A unique string/integer identifier for the image.
      detections_dict: A dictionary containing -
        DetectionResultFields.detection_scores: float32 numpy array of shape
          [num_boxes] containing detection scores for the boxes.
        DetectionResultFields.detection_classes: integer numpy array of shape
          [num_boxes] containing 1-indexed detection classes for the boxes.
        DetectionResultFields.detection_masks: optional uint8 numpy array of
          shape [num_boxes, image_height, image_width] containing instance
          masks corresponding to the boxes. The elements of the array must be
          in {0, 1}.

    Raises:
      ValueError: If groundtruth for the image_id is not available or if
        spatial shapes of groundtruth_instance_masks and detection_masks are
        incompatible.
    """
    if image_id not in self._image_id_to_mask_shape_map:
      raise ValueError('Missing groundtruth for image id: {}'.format(image_id))

    if image_id in self._image_ids_with_detections:
      tf.logging.warning('Ignoring detection with image id %s since it was '
                         'previously added', image_id)
      return

    groundtruth_masks_shape = self._image_id_to_mask_shape_map[image_id]
    detection_masks = detections_dict[standard_fields.DetectionResultFields.
                                      detection_masks]
    if groundtruth_masks_shape[1:] != detection_masks.shape[1:]:
      raise ValueError('Spatial shape of groundtruth masks and detection masks '
                       'are incompatible: {} vs {}'.format(
                           groundtruth_masks_shape,
                           detection_masks.shape))
    _check_mask_type_and_value(standard_fields.DetectionResultFields.
                               detection_masks,
                               detection_masks)
    self._detection_masks_list.extend(
        coco_tools.ExportSingleImageDetectionMasksToCoco(
            image_id=image_id,
            category_id_set=self._category_id_set,
            detection_masks=detection_masks,
            detection_scores=detections_dict[standard_fields.
                                             DetectionResultFields.
                                             detection_scores],
            detection_classes=detections_dict[standard_fields.
                                              DetectionResultFields.
                                              detection_classes]))
    self._image_ids_with_detections.update([image_id])

  def evaluate(self):
    """Evaluates the detection masks and returns a dictionary of coco metrics.

    Returns:
      A dictionary holding -

      1. summary_metrics:
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
      'DetectionMasks_Precision/mAP': mean average precision over classes
        averaged over IOU thresholds ranging from .5 to .95 with .05 increments.
      'DetectionMasks_Precision/mAP@.50IOU': mean average precision at 50% IOU.
      'DetectionMasks_Precision/mAP@.75IOU': mean average precision at 75% IOU.
      'DetectionMasks_Precision/mAP (small)': mean average precision for small
        objects (area < 32^2 pixels).
      'DetectionMasks_Precision/mAP (medium)': mean average precision for medium
        sized objects (32^2 pixels < area < 96^2 pixels).
      'DetectionMasks_Precision/mAP (large)': mean average precision for large
        objects (96^2 pixels < area < 10000^2 pixels).
      'DetectionMasks_Recall/AR@1': average recall with 1 detection.
      'DetectionMasks_Recall/AR@10': average recall with 10 detections.
      'DetectionMasks_Recall/AR@100': average recall with 100 detections.
      'DetectionMasks_Recall/AR@100 (small)': average recall for small objects
        with 100 detections.
      'DetectionMasks_Recall/AR@100 (medium)': average recall for medium objects
        with 100 detections.
      'DetectionMasks_Recall/AR@100 (large)': average recall for large objects
        with 100 detections.
527

528
529
      2. per_category_ap: if include_metrics_per_category is True, category
      specific results with keys of the form:
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
      'Precision mAP ByCategory/category' (without the supercategory part if
      no supercategories exist). For backward compatibility
      'PerformanceByCategory' is included in the output regardless of
      all_metrics_per_category.
    """
    groundtruth_dict = {
        'annotations': self._groundtruth_list,
        'images': [{'id': image_id, 'height': shape[1], 'width': shape[2]}
                   for image_id, shape in self._image_id_to_mask_shape_map.
                   iteritems()],
        'categories': self._categories
    }
    coco_wrapped_groundtruth = coco_tools.COCOWrapper(
        groundtruth_dict, detection_type='segmentation')
    coco_wrapped_detection_masks = coco_wrapped_groundtruth.LoadAnnotations(
        self._detection_masks_list)
    mask_evaluator = coco_tools.COCOEvalWrapper(
        coco_wrapped_groundtruth, coco_wrapped_detection_masks,
        agnostic_mode=False, iou_type='segm')
549
550
    mask_metrics, mask_per_category_ap = mask_evaluator.ComputeMetrics(
        include_metrics_per_category=self._include_metrics_per_category)
551
552
553
554
    mask_metrics.update(mask_per_category_ap)
    mask_metrics = {'DetectionMasks_'+ key: value
                    for key, value in mask_metrics.iteritems()}
    return mask_metrics
555
556
557
558
559

  def get_estimator_eval_metric_ops(self, image_id, groundtruth_boxes,
                                    groundtruth_classes,
                                    groundtruth_instance_masks,
                                    detection_scores, detection_classes,
560
                                    detection_masks, groundtruth_is_crowd=None):
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
    """Returns a dictionary of eval metric ops to use with `tf.EstimatorSpec`.

    Note that once value_op is called, the detections and groundtruth added via
    update_op are cleared.

    Args:
      image_id: Unique string/integer identifier for the image.
      groundtruth_boxes: float32 tensor of shape [num_boxes, 4] containing
        `num_boxes` groundtruth boxes of the format
        [ymin, xmin, ymax, xmax] in absolute image coordinates.
      groundtruth_classes: int32 tensor of shape [num_boxes] containing
        1-indexed groundtruth classes for the boxes.
      groundtruth_instance_masks: uint8 tensor array of shape
        [num_boxes, image_height, image_width] containing groundtruth masks
        corresponding to the boxes. The elements of the array must be in {0, 1}.
      detection_scores: float32 tensor of shape [num_boxes] containing
        detection scores for the boxes.
      detection_classes: int32 tensor of shape [num_boxes] containing
        1-indexed detection classes for the boxes.
      detection_masks: uint8 tensor array of shape
        [num_boxes, image_height, image_width] containing instance masks
        corresponding to the boxes. The elements of the array must be in {0, 1}.
583
584
585
      groundtruth_is_crowd: bool tensor of shape [batch, num_boxes] containing
        is_crowd annotations. This field is optional, and if not passed, then
        all boxes are treated as *not* is_crowd.
586
587
588
589
590
591
592
593
594
595
596
597

    Returns:
      a dictionary of metric names to tuple of value_op and update_op that can
      be used as eval metric ops in tf.EstimatorSpec. Note that all update ops
      must be run together and similarly all value ops must be run together to
      guarantee correct behaviour.
    """
    def update_op(
        image_id,
        groundtruth_boxes,
        groundtruth_classes,
        groundtruth_instance_masks,
598
        groundtruth_is_crowd,
599
600
601
602
603
604
605
        detection_scores,
        detection_classes,
        detection_masks):
      self.add_single_ground_truth_image_info(
          image_id,
          {'groundtruth_boxes': groundtruth_boxes,
           'groundtruth_classes': groundtruth_classes,
606
607
           'groundtruth_instance_masks': groundtruth_instance_masks,
           'groundtruth_is_crowd': groundtruth_is_crowd})
608
609
610
611
612
613
      self.add_single_detected_image_info(
          image_id,
          {'detection_scores': detection_scores,
           'detection_classes': detection_classes,
           'detection_masks': detection_masks})

614
615
    if groundtruth_is_crowd is None:
      groundtruth_is_crowd = tf.zeros_like(groundtruth_classes, dtype=tf.bool)
616
617
618
619
    update_op = tf.py_func(update_op, [image_id,
                                       groundtruth_boxes,
                                       groundtruth_classes,
                                       groundtruth_instance_masks,
620
                                       groundtruth_is_crowd,
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
                                       detection_scores,
                                       detection_classes,
                                       detection_masks], [])
    metric_names = ['DetectionMasks_Precision/mAP',
                    'DetectionMasks_Precision/mAP@.50IOU',
                    'DetectionMasks_Precision/mAP@.75IOU',
                    'DetectionMasks_Precision/mAP (large)',
                    'DetectionMasks_Precision/mAP (medium)',
                    'DetectionMasks_Precision/mAP (small)',
                    'DetectionMasks_Recall/AR@1',
                    'DetectionMasks_Recall/AR@10',
                    'DetectionMasks_Recall/AR@100',
                    'DetectionMasks_Recall/AR@100 (large)',
                    'DetectionMasks_Recall/AR@100 (medium)',
                    'DetectionMasks_Recall/AR@100 (small)']
    if self._include_metrics_per_category:
      for category_dict in self._categories:
        metric_names.append('DetectionMasks_PerformanceByCategory/mAP/' +
                            category_dict['name'])

    def first_value_func():
      self._metrics = self.evaluate()
      self.clear()
      return np.float32(self._metrics[metric_names[0]])

    def value_func_factory(metric_name):
      def value_func():
        return np.float32(self._metrics[metric_name])
      return value_func

    # Ensure that the metrics are only evaluated once.
    first_value_op = tf.py_func(first_value_func, [], tf.float32)
    eval_metric_ops = {metric_names[0]: (first_value_op, update_op)}
    with tf.control_dependencies([first_value_op]):
      for metric_name in metric_names[1:]:
        eval_metric_ops[metric_name] = (tf.py_func(
            value_func_factory(metric_name), [], np.float32), update_op)
    return eval_metric_ops