Minor fixes for object detection (tensorflow#5613)

* Internal change.

PiperOrigin-RevId: 213914693

* Add original_image_spatial_shape tensor in input dictionary to store shape of the original input image

PiperOrigin-RevId: 214018767

* Remove "groundtruth_confidences" from decoders use "groundtruth_weights" to indicate label confidence.

This also solves a bug that only surfaced now - random crop routines in core/preprocessor.py did not correctly handle "groundtruth_weight" tensors returned by the decoders.

PiperOrigin-RevId: 214091843

* Update CocoMaskEvaluator to allow for a batch of image info, rather than a single image.

PiperOrigin-RevId: 214295305

* Adding the option to be able to summarize gradients.

PiperOrigin-RevId: 214310875

* Adds FasterRCNN inference on CPU

1. Adds a flag use_static_shapes_for_eval to restrict to the ops that guarantees static shape.
2. No filtering of overlapping anchors while clipping the anchors when use_static_shapes_for_eval is set to True.
3. Adds test for faster_rcnn_meta_arch for predict and postprocess in inference mode for first and second stages.

PiperOrigin-RevId: 214329565

* Fix model_lib eval_spec_names assignment (integer->string).

PiperOrigin-RevId: 214335461

* Refactor Mask HEAD to optionally upsample after applying convolutions on ROI crops.

PiperOrigin-RevId: 214338440

* Uses final_exporter_name as exporter_name for the first eval spec for backward compatibility.

PiperOrigin-RevId: 214522032

* Add reshaped `mask_predictions` tensor to the prediction dictionary in `_predict_third_stage` method to allow computing mask loss in eval job.

PiperOrigin-RevId: 214620716

* Add support for fully conv training to fpn.

PiperOrigin-RevId: 214626274

* Fix the proprocess() function in Resnet v1 to make it work for any number of input channels.

Note: If the #channels != 3, this will simply skip the mean subtraction in preprocess() function.
PiperOrigin-RevId: 214635428

* Wrap result_dict_for_single_example in eval_util to run for batched examples.

PiperOrigin-RevId: 214678514

* Adds PNASNet-based (ImageNet model) feature extractor for SSD.

PiperOrigin-RevId: 214988331

* Update documentation

PiperOrigin-RevId: 215243502

* Correct index used to compute number of groundtruth/detection boxes in COCOMaskEvaluator.

Due to an incorrect indexing in cl/214295305 only the first detection mask and first groundtruth mask for a given image are fed to the COCO Mask evaluation library. Since groundtruth masks are arranged in no particular order, the first and highest scoring detection mask (detection masks are ordered by score) won't match the the first and only groundtruth retained in all cases. This is I think why mask evaluation metrics do not get better than ~11 mAP. Note that this code path is only active when using model_main.py binary for evaluation.

This change fixes the indices and modifies an existing test case to cover it.

PiperOrigin-RevId: 215275936

* Fixing grayscale_image_resizer to accept mask as input.

PiperOrigin-RevId: 215345836

* Add an option not to clip groundtruth boxes during preprocessing. Clipping boxes adversely affects training for partially occluded or large objects, especially for fully conv models. Clipping already occurs during postprocessing, and should not occur during training.

PiperOrigin-RevId: 215613379

* Always return recalls and precisions with length equal to the number of classes.

The previous behavior of ObjectDetectionEvaluation was somewhat dangerous: when no groundtruth boxes were present, the lists of per-class precisions and recalls were simply truncated. Unless you were aware of this phenomenon (and consulted the `num_gt_instances_per_class` vector) it was difficult to associate each metric with each class.

PiperOrigin-RevId: 215633711

* Expose the box feature node in SSD.

PiperOrigin-RevId: 215653316

* Fix ssd mobilenet v2 _CONV_DEFS overwriting issue.

PiperOrigin-RevId: 215654160

* More documentation updates

PiperOrigin-RevId: 215656580

* Add pooling + residual option in multi_resolution_feature_maps. It adds an average pooling and a residual layer between feature maps with matching depth. Designed to be used with WeightSharedBoxPredictor.

PiperOrigin-RevId: 215665619

* Only call create_modificed_mobilenet_config on init if use_depthwise is true.

PiperOrigin-RevId: 215784290

* Only call create_modificed_mobilenet_config on init if use_depthwise is true.

PiperOrigin-RevId: 215837524

* Don't prune keypoints if clip_boxes is false.

PiperOrigin-RevId: 216187642

* Makes sure "key" field exists in the result dictionary.

PiperOrigin-RevId: 216456543

* Add add_background_class parameter to allow disabling the inclusion of a background class.

PiperOrigin-RevId: 216567612

* Update expected_classification_loss_under_sampling to better account for expected sampling.

PiperOrigin-RevId: 216712287

* Let the evaluation receive a evaluation class in its constructor.

PiperOrigin-RevId: 216769374

* This CL adds model building & training support for end-to-end Keras-based SSD models. If a Keras feature extractor's name is specified in the model config (e.g. 'ssd_mobilenet_v2_keras'), the model will use that feature extractor and a corresponding Keras-based box predictor.

This CL makes sure regularization losses & batch norm updates work correctly when training models that have Keras-based components. It also updates the default hyperparameter settings of the keras-based mobilenetV2 (when not overriding hyperparams) to more closely match the legacy Slim training scope.

PiperOrigin-RevId: 216938707

* Adding the ability in the coco evaluator to indicate whether an image has been annotated. For a non-annotated image, detections and groundtruth are not supplied.

PiperOrigin-RevId: 217316342

* Release the 8k minival dataset ids for MSCOCO, used in Huang et al. "Speed/accuracy trade-offs for modern convolutional object detectors" (https://arxiv.org/abs/1611.10012)

PiperOrigin-RevId: 217549353

* Exposes weighted_sigmoid_focal loss for faster rcnn classifier

PiperOrigin-RevId: 217601740

* Add detection_features to output nodes. The shape of the feature is [batch_size, max_detections, depth].

PiperOrigin-RevId: 217629905

* FPN uses a custom NN resize op for TPU-compatibility. Replace this op with the Tensorflow version at export time for TFLite-compatibility.

PiperOrigin-RevId: 217721184

* Compute `num_groundtruth_boxes` in inputs.tranform_input_data_fn after data augmentation instead of decoders.

PiperOrigin-RevId: 217733432

* 1. Stop gradients from flowing into groundtruth masks with zero paddings.
2. Normalize pixelwise cross entropy loss across the whole batch.

PiperOrigin-RevId: 217735114

* Optimize Input pipeline for Mask R-CNN on TPU with blfoat16: improve the step time from:
1663.6 ms -> 1184.2 ms, about 28.8% improvement.

PiperOrigin-RevId: 217748833

* Fixes to export a TPU compatible model

Adds nodes to each of the output tensor. Also increments the value of class labels by 1.

PiperOrigin-RevId: 217856760

* API changes:
 - change the interface of target assigner to return per-class weights.
 - change the interface of classification loss to take per-class weights.

PiperOrigin-RevId: 217968393

* Add an option to override pipeline config in export_saved_model using command line arg

PiperOrigin-RevId: 218429292

* Include Quantized trained MobileNet V2 SSD and FaceSsd in model zoo.

PiperOrigin-RevId: 218530947

* Write final config to disk in `train` mode only.

PiperOrigin-RevId: 218735512

research/object_detection/anchor_generators/multiscale_grid_anchor_generator.py

@@ -108,9 +108,6 @@ def _generate(self, feature_map_shape_list, im_height=1, im_width=1):
       ValueError: if im_height and im_width are 1, but normalized coordinates
         were requested.
     """
-    if not isinstance(im_height, int) or not isinstance(im_width, int):
-      raise ValueError('MultiscaleGridAnchorGenerator currently requires '
-                       'input image shape to be statically defined.')
     anchor_grid_list = []
     for feat_shape, grid_info in zip(feature_map_shape_list,
                                      self._anchor_grid_info):
@@ -122,10 +119,11 @@ def _generate(self, feature_map_shape_list, im_height=1, im_width=1):
       feat_h = feat_shape[0]
       feat_w = feat_shape[1]
       anchor_offset = [0, 0]
-      if im_height % 2.0**level == 0 or im_height == 1:
-        anchor_offset[0] = stride / 2.0
-      if im_width % 2.0**level == 0 or im_width == 1:
-        anchor_offset[1] = stride / 2.0
+      if isinstance(im_height, int) and isinstance(im_width, int):
+        if im_height % 2.0**level == 0 or im_height == 1:
+          anchor_offset[0] = stride / 2.0
+        if im_width % 2.0**level == 0 or im_width == 1:
+          anchor_offset[1] = stride / 2.0
       ag = grid_anchor_generator.GridAnchorGenerator(
           scales,
           aspect_ratios,

research/object_detection/anchor_generators/multiscale_grid_anchor_generator_test.py

@@ -116,7 +116,7 @@ def test_num_anchors_per_location(self):
         normalize_coordinates=False)
     self.assertEqual(anchor_generator.num_anchors_per_location(), [6, 6])
 
-  def test_construct_single_anchor_fails_with_tensor_image_size(self):
+  def test_construct_single_anchor_dynamic_size(self):
     min_level = 5
     max_level = 5
     anchor_scale = 4.0
@@ -125,12 +125,22 @@ def test_construct_single_anchor_fails_with_tensor_image_size(self):
     im_height = tf.constant(64)
     im_width = tf.constant(64)
     feature_map_shape_list = [(2, 2)]
+    # Zero offsets are used.
+    exp_anchor_corners = [[-64, -64, 64, 64],
+                          [-64, -32, 64, 96],
+                          [-32, -64, 96, 64],
+                          [-32, -32, 96, 96]]
+
     anchor_generator = mg.MultiscaleGridAnchorGenerator(
         min_level, max_level, anchor_scale, aspect_ratios, scales_per_octave,
         normalize_coordinates=False)
-    with self.assertRaisesRegexp(ValueError, 'statically defined'):
-      anchor_generator.generate(
-          feature_map_shape_list, im_height=im_height, im_width=im_width)
+    anchors_list = anchor_generator.generate(
+        feature_map_shape_list, im_height=im_height, im_width=im_width)
+    anchor_corners = anchors_list[0].get()
+
+    with self.test_session():
+      anchor_corners_out = anchor_corners.eval()
+      self.assertAllClose(anchor_corners_out, exp_anchor_corners)
 
   def test_construct_single_anchor_with_odd_input_dimension(self):
 

research/object_detection/builders/box_predictor_builder_test.py

@@ -113,7 +113,8 @@ def mock_conv_argscope_builder(conv_hyperparams_arg, is_training):
         argscope_fn=mock_conv_argscope_builder,
         box_predictor_config=box_predictor_proto,
         is_training=False,
-        num_classes=10)
+        num_classes=10,
+        add_background_class=False)
     class_head = box_predictor._class_prediction_head
     self.assertEqual(box_predictor._min_depth, 2)
     self.assertEqual(box_predictor._max_depth, 16)
@@ -122,6 +123,7 @@ def mock_conv_argscope_builder(conv_hyperparams_arg, is_training):
     self.assertAlmostEqual(class_head._dropout_keep_prob, 0.4)
     self.assertTrue(class_head._apply_sigmoid_to_scores)
     self.assertAlmostEqual(class_head._class_prediction_bias_init, 4.0)
+    self.assertEqual(class_head._num_class_slots, 10)
     self.assertEqual(box_predictor.num_classes, 10)
     self.assertFalse(box_predictor._is_training)
     self.assertTrue(class_head._use_depthwise)
@@ -154,6 +156,7 @@ def test_construct_default_conv_box_predictor(self):
     self.assertTrue(class_head._use_dropout)
     self.assertAlmostEqual(class_head._dropout_keep_prob, 0.8)
     self.assertFalse(class_head._apply_sigmoid_to_scores)
+    self.assertEqual(class_head._num_class_slots, 91)
     self.assertEqual(box_predictor.num_classes, 90)
     self.assertTrue(box_predictor._is_training)
     self.assertFalse(class_head._use_depthwise)
@@ -306,7 +309,8 @@ def mock_conv_argscope_builder(conv_hyperparams_arg, is_training):
         argscope_fn=mock_conv_argscope_builder,
         box_predictor_config=box_predictor_proto,
         is_training=False,
-        num_classes=10)
+        num_classes=10,
+        add_background_class=False)
     class_head = box_predictor._class_prediction_head
     self.assertEqual(box_predictor._depth, 2)
     self.assertEqual(box_predictor._num_layers_before_predictor, 2)
@@ -349,7 +353,8 @@ def mock_conv_argscope_builder(conv_hyperparams_arg, is_training):
         argscope_fn=mock_conv_argscope_builder,
         box_predictor_config=box_predictor_proto,
         is_training=False,
-        num_classes=10)
+        num_classes=10,
+        add_background_class=False)
     class_head = box_predictor._class_prediction_head
     self.assertEqual(box_predictor._depth, 2)
     self.assertEqual(box_predictor._num_layers_before_predictor, 2)
@@ -627,6 +632,48 @@ def test_build_box_predictor_with_mask_branch(self):
         third_stage_heads[mask_rcnn_box_predictor.MASK_PREDICTIONS]
         ._mask_prediction_conv_depth, 512)
 
+  def test_build_box_predictor_with_convlve_then_upsample_masks(self):
+    box_predictor_proto = box_predictor_pb2.BoxPredictor()
+    box_predictor_proto.mask_rcnn_box_predictor.fc_hyperparams.op = (
+        hyperparams_pb2.Hyperparams.FC)
+    box_predictor_proto.mask_rcnn_box_predictor.conv_hyperparams.op = (
+        hyperparams_pb2.Hyperparams.CONV)
+    box_predictor_proto.mask_rcnn_box_predictor.predict_instance_masks = True
+    box_predictor_proto.mask_rcnn_box_predictor.mask_prediction_conv_depth = 512
+    box_predictor_proto.mask_rcnn_box_predictor.mask_height = 24
+    box_predictor_proto.mask_rcnn_box_predictor.mask_width = 24
+    box_predictor_proto.mask_rcnn_box_predictor.convolve_then_upsample_masks = (
+        True)
+
+    mock_argscope_fn = mock.Mock(return_value='arg_scope')
+    box_predictor = box_predictor_builder.build(
+        argscope_fn=mock_argscope_fn,
+        box_predictor_config=box_predictor_proto,
+        is_training=True,
+        num_classes=90)
+    mock_argscope_fn.assert_has_calls(
+        [mock.call(box_predictor_proto.mask_rcnn_box_predictor.fc_hyperparams,
+                   True),
+         mock.call(box_predictor_proto.mask_rcnn_box_predictor.conv_hyperparams,
+                   True)], any_order=True)
+    box_head = box_predictor._box_prediction_head
+    class_head = box_predictor._class_prediction_head
+    third_stage_heads = box_predictor._third_stage_heads
+    self.assertFalse(box_head._use_dropout)
+    self.assertFalse(class_head._use_dropout)
+    self.assertAlmostEqual(box_head._dropout_keep_prob, 0.5)
+    self.assertAlmostEqual(class_head._dropout_keep_prob, 0.5)
+    self.assertEqual(box_predictor.num_classes, 90)
+    self.assertTrue(box_predictor._is_training)
+    self.assertEqual(box_head._box_code_size, 4)
+    self.assertTrue(
+        mask_rcnn_box_predictor.MASK_PREDICTIONS in third_stage_heads)
+    self.assertEqual(
+        third_stage_heads[mask_rcnn_box_predictor.MASK_PREDICTIONS]
+        ._mask_prediction_conv_depth, 512)
+    self.assertTrue(third_stage_heads[mask_rcnn_box_predictor.MASK_PREDICTIONS]
+                    ._convolve_then_upsample)
+
 
 class RfcnBoxPredictorBuilderTest(tf.test.TestCase):
 

research/object_detection/builders/hyperparams_builder.py

@@ -64,6 +64,10 @@ def __init__(self, hyperparams_config):
           hyperparams_config.batch_norm)
 
     self._activation_fn = _build_activation_fn(hyperparams_config.activation)
+    # TODO(kaftan): Unclear if these kwargs apply to separable & depthwise conv
+    # (Those might use depthwise_* instead of kernel_*)
+    # We should probably switch to using build_conv2d_layer and
+    # build_depthwise_conv2d_layer methods instead.
     self._op_params = {
         'kernel_regularizer': _build_keras_regularizer(
             hyperparams_config.regularizer),

research/object_detection/builders/image_resizer_builder.py

@@ -106,10 +106,35 @@ def build(image_resizer_config):
     raise ValueError(
         'Invalid image resizer option: \'%s\'.' % image_resizer_oneof)
 
-  def grayscale_image_resizer(image):
-    [resized_image, resized_image_shape] = image_resizer_fn(image)
-    grayscale_image = preprocessor.rgb_to_gray(resized_image)
-    grayscale_image_shape = tf.concat([resized_image_shape[:-1], [1]], 0)
-    return [grayscale_image, grayscale_image_shape]
+  def grayscale_image_resizer(image, masks=None):
+    """Convert to grayscale before applying image_resizer_fn.
+
+    Args:
+      image: A 3D tensor of shape [height, width, 3]
+      masks: (optional) rank 3 float32 tensor with shape [num_instances, height,
+        width] containing instance masks.
+
+    Returns:
+    Note that the position of the resized_image_shape changes based on whether
+    masks are present.
+    resized_image: A 3D tensor of shape [new_height, new_width, 1],
+      where the image has been resized (with bilinear interpolation) so that
+      min(new_height, new_width) == min_dimension or
+      max(new_height, new_width) == max_dimension.
+    resized_masks: If masks is not None, also outputs masks. A 3D tensor of
+      shape [num_instances, new_height, new_width].
+    resized_image_shape: A 1D tensor of shape [3] containing shape of the
+      resized image.
+    """
+    # image_resizer_fn returns [resized_image, resized_image_shape] if
+    # mask==None, otherwise it returns
+    # [resized_image, resized_mask, resized_image_shape]. In either case, we
+    # only deal with first and last element of the returned list.
+    retval = image_resizer_fn(image, masks)
+    resized_image = retval[0]
+    resized_image_shape = retval[-1]
+    retval[0] = preprocessor.rgb_to_gray(resized_image)
+    retval[-1] = tf.concat([resized_image_shape[:-1], [1]], 0)
+    return retval
 
   return functools.partial(grayscale_image_resizer)

research/object_detection/builders/losses_builder.py

@@ -136,6 +136,14 @@ def build_faster_rcnn_classification_loss(loss_config):
     config = loss_config.weighted_logits_softmax
     return losses.WeightedSoftmaxClassificationAgainstLogitsLoss(
         logit_scale=config.logit_scale)
+  if loss_type == 'weighted_sigmoid_focal':
+    config = loss_config.weighted_sigmoid_focal
+    alpha = None
+    if config.HasField('alpha'):
+      alpha = config.alpha
+    return losses.SigmoidFocalClassificationLoss(
+        gamma=config.gamma,
+        alpha=alpha)
 
   # By default, Faster RCNN second stage classifier uses Softmax loss
   # with anchor-wise outputs.

research/object_detection/builders/losses_builder_test.py

@@ -280,7 +280,7 @@ def test_anchorwise_output(self):
                                losses.WeightedSigmoidClassificationLoss))
     predictions = tf.constant([[[0.0, 1.0, 0.0], [0.0, 0.5, 0.5]]])
     targets = tf.constant([[[0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]])
-    weights = tf.constant([[1.0, 1.0]])
+    weights = tf.constant([[[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]])
     loss = classification_loss(predictions, targets, weights=weights)
     self.assertEqual(loss.shape, [1, 2, 3])
 
@@ -473,6 +473,19 @@ def test_build_logits_softmax_loss(self):
         isinstance(classification_loss,
                    losses.WeightedSoftmaxClassificationAgainstLogitsLoss))
 
+  def test_build_sigmoid_focal_loss(self):
+    losses_text_proto = """
+      weighted_sigmoid_focal {
+      }
+    """
+    losses_proto = losses_pb2.ClassificationLoss()
+    text_format.Merge(losses_text_proto, losses_proto)
+    classification_loss = losses_builder.build_faster_rcnn_classification_loss(
+        losses_proto)
+    self.assertTrue(
+        isinstance(classification_loss,
+                   losses.SigmoidFocalClassificationLoss))
+
   def test_build_softmax_loss_by_default(self):
     losses_text_proto = """
     """