def create_target_assigner(reference, stage=None,
negative_class_weight=1.0,
unmatched_cls_target=None):
"""Factory function for creating standard target assigners.
Args:
reference: string referencing the type of TargetAssigner.
stage: string denoting stage: {proposal, detection}.
negative_class_weight: classification weight to be associated to negative
anchors (default: 1.0)
unmatched_cls_target: a float32 tensor with shape [d_1, d_2, ..., d_k]
which is consistent with the classification target for each
anchor (and can be empty for scalar targets). This shape must thus be
compatible with the groundtruth labels that are passed to the Assign
function (which have shape [num_gt_boxes, d_1, d_2, ..., d_k]).
If set to None, unmatched_cls_target is set to be 0 for each anchor.
Returns:
TargetAssigner: desired target assigner.
Raises:
ValueError: if combination reference+stage is invalid.
"""
if reference == 'Multibox' and stage == 'proposal':
similarity_calc = sim_calc.NegSqDistSimilarity()
matcher = bipartite_matcher.GreedyBipartiteMatcher()
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder()
elif reference == 'FasterRCNN' and stage == 'proposal':
similarity_calc = sim_calc.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.7,
unmatched_threshold=0.3,
force_match_for_each_row=True)
box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder(
scale_factors=[10.0, 10.0, 5.0, 5.0])
elif reference == 'FasterRCNN' and stage == 'detection':
similarity_calc = sim_calc.IouSimilarity()
# Uses all proposals with IOU < 0.5 as candidate negatives.
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
negatives_lower_than_unmatched=True)
box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder(
scale_factors=[10.0, 10.0, 5.0, 5.0])
elif reference == 'FastRCNN':
similarity_calc = sim_calc.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.1,
force_match_for_each_row=False,
negatives_lower_than_unmatched=False)
box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder()
else:
raise ValueError('No valid combination of reference and stage.')
return TargetAssigner(similarity_calc, matcher, box_coder,
negative_class_weight=negative_class_weight,
unmatched_cls_target=unmatched_cls_target)
def create_target_assigner(reference, stage=None,
negative_class_weight=1.0, use_matmul_gather=False):
"""Factory function for creating standard target assigners.
Args:
reference: string referencing the type of TargetAssigner.
stage: string denoting stage: {proposal, detection}.
negative_class_weight: classification weight to be associated to negative
anchors (default: 1.0)
use_matmul_gather: whether to use matrix multiplication based gather which
are better suited for TPUs.
Returns:
TargetAssigner: desired target assigner.
Raises:
ValueError: if combination reference+stage is invalid.
"""
if reference == 'Multibox' and stage == 'proposal':
similarity_calc = sim_calc.NegSqDistSimilarity()
matcher = bipartite_matcher.GreedyBipartiteMatcher()
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder()
elif reference == 'FasterRCNN' and stage == 'proposal':
similarity_calc = sim_calc.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.7,
unmatched_threshold=0.3,
force_match_for_each_row=True,
use_matmul_gather=use_matmul_gather)
box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder(
scale_factors=[10.0, 10.0, 5.0, 5.0])
elif reference == 'FasterRCNN' and stage == 'detection':
similarity_calc = sim_calc.IouSimilarity()
# Uses all proposals with IOU < 0.5 as candidate negatives.
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
negatives_lower_than_unmatched=True,
use_matmul_gather=use_matmul_gather)
box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder(
scale_factors=[10.0, 10.0, 5.0, 5.0])
elif reference == 'FastRCNN':
similarity_calc = sim_calc.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.1,
force_match_for_each_row=False,
negatives_lower_than_unmatched=False,
use_matmul_gather=use_matmul_gather)
box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder()
else:
raise ValueError('No valid combination of reference and stage.')
return TargetAssigner(similarity_calc, matcher, box_coder,
negative_class_weight=negative_class_weight)
def build(region_similarity_calculator_config):
"""Builds region similarity calculator based on the configuration.
Builds one of [IouSimilarity, IoaSimilarity, NegSqDistSimilarity] objects. See
core/region_similarity_calculator.proto for details.
Args:
region_similarity_calculator_config: RegionSimilarityCalculator
configuration proto.
Returns:
region_similarity_calculator: RegionSimilarityCalculator object.
Raises:
ValueError: On unknown region similarity calculator.
"""
if not isinstance(
region_similarity_calculator_config,
region_similarity_calculator_pb2.RegionSimilarityCalculator):
raise ValueError(
'region_similarity_calculator_config not of type '
'region_similarity_calculator_pb2.RegionsSimilarityCalculator')
similarity_calculator = region_similarity_calculator_config.WhichOneof(
'region_similarity')
if similarity_calculator == 'iou_similarity':
return region_similarity_calculator.IouSimilarity()
if similarity_calculator == 'ioa_similarity':
return region_similarity_calculator.IoaSimilarity()
if similarity_calculator == 'neg_sq_dist_similarity':
return region_similarity_calculator.NegSqDistSimilarity()
raise ValueError('Unknown region similarity calculator.')
def _get_agnostic_target_assigner(self):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.5)
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder()
return targetassigner.TargetAssigner(
similarity_calc, matcher, box_coder,
unmatched_cls_target=None)
def _get_multi_class_target_assigner(self, num_classes):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.5)
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder()
unmatched_cls_target = tf.constant([1] + num_classes * [0], tf.float32)
return targetassigner.TargetAssigner(
similarity_calc, matcher, box_coder,
unmatched_cls_target=unmatched_cls_target)
def _get_multi_dimensional_target_assigner(self, target_dimensions):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.5)
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder()
unmatched_cls_target = tf.constant(np.zeros(target_dimensions),
tf.float32)
return targetassigner.TargetAssigner(
similarity_calc, matcher, box_coder,
unmatched_cls_target=unmatched_cls_target)
def _create_model(self, apply_hard_mining=True):
is_training = False
num_classes = 1
mock_anchor_generator = MockAnchorGenerator2x2()
mock_box_predictor = test_utils.MockBoxPredictor(
is_training, num_classes)
mock_box_coder = test_utils.MockBoxCoder()
fake_feature_extractor = FakeSSDFeatureExtractor()
mock_matcher = test_utils.MockMatcher()
region_similarity_calculator = sim_calc.IouSimilarity()
encode_background_as_zeros = False
def image_resizer_fn(image):
return [tf.identity(image), tf.shape(image)]
classification_loss = losses.WeightedSigmoidClassificationLoss()
localization_loss = losses.WeightedSmoothL1LocalizationLoss()
non_max_suppression_fn = functools.partial(
post_processing.batch_multiclass_non_max_suppression,
score_thresh=-20.0,
iou_thresh=1.0,
max_size_per_class=5,
max_total_size=5)
classification_loss_weight = 1.0
localization_loss_weight = 1.0
normalize_loss_by_num_matches = False
hard_example_miner = None
if apply_hard_mining:
# This hard example miner is expected to be a no-op.
hard_example_miner = losses.HardExampleMiner(
num_hard_examples=None, iou_threshold=1.0)
code_size = 4
model = ssd_meta_arch.SSDMetaArch(is_training,
mock_anchor_generator,
mock_box_predictor,
mock_box_coder,
fake_feature_extractor,
mock_matcher,
region_similarity_calculator,
encode_background_as_zeros,
image_resizer_fn,
non_max_suppression_fn,
tf.identity,
classification_loss,
localization_loss,
classification_loss_weight,
localization_loss_weight,
normalize_loss_by_num_matches,
hard_example_miner,
add_summaries=False)
return model, num_classes, mock_anchor_generator.num_anchors(
), code_size
def test_get_correct_pairwise_similarity_based_on_iou(self):
corners1 = tf.constant([[4.0, 3.0, 7.0, 5.0], [5.0, 6.0, 10.0, 7.0]])
corners2 = tf.constant([[3.0, 4.0, 6.0, 8.0], [14.0, 14.0, 15.0, 15.0],
[0.0, 0.0, 20.0, 20.0]])
exp_output = [[2.0 / 16.0, 0, 6.0 / 400.0], [1.0 / 16.0, 0.0, 5.0 / 400.0]]
boxes1 = box_list.BoxList(corners1)
boxes2 = box_list.BoxList(corners2)
iou_similarity_calculator = region_similarity_calculator.IouSimilarity()
iou_similarity = iou_similarity_calculator.compare(boxes1, boxes2)
with self.test_session() as sess:
iou_output = sess.run(iou_similarity)
self.assertAllClose(iou_output, exp_output)
def test_get_correct_pairwise_similarity_based_on_iou_rbox(self):
corners1 = tf.constant([[4.0, 3.0, 7.0, 5.0, 0.0], [5.0, 6.0, 10.0, 7.0, 0.0]])
corners2 = tf.constant([[3.0, 4.0, 6.0, 8.0, 0.0], [14.0, 14.0, 15.0, 15.0, 0.0],
[0.0, 0.0, 20.0, 20.0, 0.0]])
exp_output = [[0.495495, 0, 0.0875], [0.388235, 0.036907, 0.175]]
rboxes1 = rbox_list.RBoxList(corners1)
rboxes2 = rbox_list.RBoxList(corners2)
iou_similarity_calculator = region_similarity_calculator.IouSimilarity()
iou_similarity = iou_similarity_calculator.compare(rboxes1, rboxes2)
with self.test_session() as sess:
iou_output = sess.run(iou_similarity)
self.assertAllClose(iou_output, exp_output)
def test_assign_multidimensional_class_targets_rbox(self):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.3)
box_coder = faster_rcnn_rbox_coder.FasterRcnnRBoxCoder()
unmatched_cls_target = tf.constant([[0, 0], [0, 0]], tf.float32)
target_assigner = targetassigner.TargetAssigner(
similarity_calc,
matcher,
box_coder,
unmatched_cls_target=unmatched_cls_target)
anchors = tf.constant([[0.0, 0.0, 0.5, 0.5, 0.0],
[0.5, 0.5, 1.0, 0.8,
0.0], [0, 0.5, .5, 1.0, 0.0],
[.75, 0, 1.0, .25, 0.0]])
anchors_rbox = rbox_list.RBoxList(anchors)
box_corners = [[0.0, 0.0, 0.5, 0.5, 0.0], [0.5, 0.5, 0.9, 0.9, 0.0],
[.75, 0, .95, .27, 0.0]]
rboxes = rbox_list.RBoxList(tf.constant(box_corners))
groundtruth_labels = tf.constant(
[[[0, 1], [1, 0]], [[1, 0], [0, 1]], [[0, 1], [1, .5]]],
tf.float32)
exp_cls_targets = [[[0, 1], [1, 0]], [[1, 0], [0, 1]], [[0, 0], [0,
0]],
[[0, 1], [1, .5]]]
exp_cls_weights = [1, 1, 1, 1]
exp_reg_targets = [[0, 0, 0, 0, 0], [0, 0, -0.105361, 0.117783, 0],
[0, 0, 0, 0, 0], [0, 0, -0.0512933, .0769611, 0]]
exp_reg_weights = [1, 1, 0, 1]
exp_matching_anchors = [0, 1, 3]
result = target_assigner.assign(anchors_rbox, rboxes,
groundtruth_labels)
(cls_targets, cls_weights, reg_targets, reg_weights, match) = result
with self.test_session() as sess:
(cls_targets_out, cls_weights_out, reg_targets_out, reg_weights_out, matching_anchors_out) = \
sess.run([cls_targets, cls_weights, reg_targets, reg_weights, match.matched_column_indices()])
self.assertAllClose(cls_targets_out, exp_cls_targets)
self.assertAllClose(cls_weights_out, exp_cls_weights)
self.assertAllClose(reg_targets_out, exp_reg_targets)
self.assertAllClose(reg_weights_out, exp_reg_weights)
self.assertAllClose(matching_anchors_out, exp_matching_anchors)
self.assertEquals(cls_targets_out.dtype, np.float32)
self.assertEquals(cls_weights_out.dtype, np.float32)
self.assertEquals(reg_targets_out.dtype, np.float32)
self.assertEquals(reg_weights_out.dtype, np.float32)
self.assertEquals(matching_anchors_out.dtype, np.int32)
def graph_fn(anchor_means, anchor_stddevs, groundtruth_box_corners):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.3)
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder()
target_assigner = targetassigner.TargetAssigner(
similarity_calc, matcher, box_coder, unmatched_cls_target=None)
anchors_boxlist = box_list.BoxList(anchor_means)
anchors_boxlist.add_field('stddev', anchor_stddevs)
groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
result = target_assigner.assign(anchors_boxlist, groundtruth_boxlist)
(cls_targets, cls_weights, reg_targets, reg_weights, _) = result
return (cls_targets, cls_weights, reg_targets, reg_weights)
def setUp(self):
"""Set up mock RSSD model.
Here we set up a simple mock RSSD model that will always predict 4
detections that happen to always be exactly the anchors that are set up
in the above MockAnchorGenerator. Because we let max_detections=5,
we will also always end up with an extra padded row in the detection
results.
"""
is_training = False
self._num_classes = 1
mock_anchor_generator = MockAnchorGenerator2x2()
mock_rbox_predictor = test_utils.MockRBoxPredictor(
is_training, self._num_classes)
mock_rbox_coder = test_utils.MockRBoxCoder()
fake_feature_extractor = FakeSSDFeatureExtractor()
mock_matcher = test_utils.MockMatcher()
region_similarity_calculator = sim_calc.IouSimilarity()
def image_resizer_fn(image):
return tf.identity(image)
classification_loss = losses.WeightedSigmoidClassificationLoss(
anchorwise_output=True)
localization_loss = losses.WeightedSmoothL1LocalizationLoss(
anchorwise_output=True)
non_max_suppression_fn = functools.partial(
post_processing_rbox.batch_multiclass_non_max_suppression_rbox,
score_thresh=-20.0,
iou_thresh=1.0,
max_size_per_class=5,
max_total_size=5)
classification_loss_weight = 1.0
localization_loss_weight = 1.0
normalize_loss_by_num_matches = False
# This hard example miner is expected to be a no-op.
hard_example_miner = losses.HardExampleMiner(num_hard_examples=None,
iou_threshold=1.0,
box_type='rbbox')
self._num_anchors = 4
self._code_size = 5
self._model = rssd_meta_arch.RSSDMetaArch(
is_training, mock_anchor_generator, mock_rbox_predictor,
mock_rbox_coder, fake_feature_extractor, mock_matcher,
region_similarity_calculator, image_resizer_fn,
non_max_suppression_fn, tf.identity, classification_loss,
localization_loss, classification_loss_weight,
localization_loss_weight, normalize_loss_by_num_matches,
hard_example_miner)
def _get_multi_dimensional_target_assigner_rbox(self, target_dimensions):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.3)
box_coder = faster_rcnn_rbox_coder.FasterRcnnRBoxCoder()
unmatched_cls_target = tf.constant(np.zeros(target_dimensions),
tf.float32)
return targetassigner.TargetAssigner(
similarity_calc,
matcher,
box_coder,
positive_class_weight=1.0,
negative_class_weight=1.0,
unmatched_cls_target=unmatched_cls_target)
def graph_fn(anchor_means, groundtruth_box_corners,
groundtruth_keypoints):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.5)
box_coder = keypoint_box_coder.KeypointBoxCoder(
num_keypoints=6, scale_factors=[10.0, 10.0, 5.0, 5.0])
target_assigner = targetassigner.TargetAssigner(
similarity_calc, matcher, box_coder, unmatched_cls_target=None)
anchors_boxlist = box_list.BoxList(anchor_means)
groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
groundtruth_boxlist.add_field(fields.BoxListFields.keypoints,
groundtruth_keypoints)
result = target_assigner.assign(anchors_boxlist, groundtruth_boxlist)
(cls_targets, cls_weights, reg_targets, reg_weights, _) = result
return (cls_targets, cls_weights, reg_targets, reg_weights)
def test_assign_with_ignored_matches(self):
# Note: test is very similar to above. The third box matched with an IOU
# of 0.35, which is between the matched and unmatched threshold. This means
# That like above the expected classification targets are [1, 1, 0].
# Unlike above, the third target is ignored and therefore expected
# classification weights are [1, 1, 0].
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.3)
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder()
target_assigner = targetassigner.TargetAssigner(
similarity_calc, matcher, box_coder)
prior_means = tf.constant([[0.0, 0.0, 0.5, 0.5], [0.5, 0.5, 1.0, 0.8],
[0.0, 0.5, .9, 1.0]])
prior_stddevs = tf.constant(3 * [4 * [.1]])
priors = box_list.BoxList(prior_means)
priors.add_field('stddev', prior_stddevs)
box_corners = [[0.0, 0.0, 0.5, 0.5], [0.5, 0.5, 0.9, 0.9]]
boxes = box_list.BoxList(tf.constant(box_corners))
exp_cls_targets = [[1], [1], [0]]
exp_cls_weights = [1, 1, 0]
exp_reg_targets = [[0, 0, 0, 0], [0, 0, -1, 1], [0, 0, 0, 0]]
exp_reg_weights = [1, 1, 0]
exp_matching_anchors = [0, 1]
result = target_assigner.assign(priors, boxes)
(cls_targets, cls_weights, reg_targets, reg_weights, match) = result
with self.test_session() as sess:
(cls_targets_out, cls_weights_out, reg_targets_out,
reg_weights_out, matching_anchors_out) = sess.run([
cls_targets, cls_weights, reg_targets, reg_weights,
match.matched_column_indices()
])
self.assertAllClose(cls_targets_out, exp_cls_targets)
self.assertAllClose(cls_weights_out, exp_cls_weights)
self.assertAllClose(reg_targets_out, exp_reg_targets)
self.assertAllClose(reg_weights_out, exp_reg_weights)
self.assertAllClose(matching_anchors_out, exp_matching_anchors)
self.assertEquals(cls_targets_out.dtype, np.float32)
self.assertEquals(cls_weights_out.dtype, np.float32)
self.assertEquals(reg_targets_out.dtype, np.float32)
self.assertEquals(reg_weights_out.dtype, np.float32)
self.assertEquals(matching_anchors_out.dtype, np.int32)
def __init__(self, categories, iou_threshold=0.5):
"""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'.
iou_threshold: Threshold above which to consider a box as matched during
evaluation.
"""
super(CalibrationDetectionEvaluator, self).__init__(categories)
# Constructing target_assigner to match detections to groundtruth.
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(
matched_threshold=iou_threshold, unmatched_threshold=iou_threshold)
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
self._target_assigner = target_assigner.TargetAssigner(
similarity_calc, matcher, box_coder)
