Implement embedding based similarity mask head for DeepMAC.

PiperOrigin-RevId: 394512536

research/object_detection/builders/losses_builder.py

@@ -263,7 +263,8 @@ def _build_classification_loss(loss_config):
   elif loss_type == 'weighted_dice_classification_loss':
     config = loss_config.weighted_dice_classification_loss
     return losses.WeightedDiceClassificationLoss(
-        squared_normalization=config.squared_normalization)
+        squared_normalization=config.squared_normalization,
+        is_prediction_probability=config.is_prediction_probability)
 
   else:
     raise ValueError('Empty loss config.')

research/object_detection/core/losses.py

@@ -286,15 +286,19 @@ class WeightedDiceClassificationLoss(Loss):
 
   """
 
-  def __init__(self, squared_normalization):
+  def __init__(self, squared_normalization, is_prediction_probability=False):
     """Initializes the loss object.
 
     Args:
       squared_normalization: boolean, if set, we square the probabilities in the
         denominator term used for normalization.
+      is_prediction_probability: boolean, whether or not the input
+        prediction_tensor represents a probability. If false, it is
+        first converted to a probability by applying sigmoid.
     """
 
     self._squared_normalization = squared_normalization
+    self.is_prediction_probability = is_prediction_probability
     super(WeightedDiceClassificationLoss, self).__init__()
 
   def _compute_loss(self,
@@ -332,6 +336,9 @@ class WeightedDiceClassificationLoss(Loss):
                                       tf.shape(prediction_tensor)[2]),
           [1, 1, -1])
 
+    if self.is_prediction_probability:
+      prob_tensor = prediction_tensor
+    else:
       prob_tensor = tf.nn.sigmoid(prediction_tensor)
 
     if self._squared_normalization:

research/object_detection/meta_architectures/deepmac_meta_arch.py

@@ -36,7 +36,8 @@ class DeepMACParams(
         'allowed_masked_classes_ids', 'mask_size', 'mask_num_subsamples',
         'use_xy', 'network_type', 'use_instance_embedding', 'num_init_channels',
         'predict_full_resolution_masks', 'postprocess_crop_size',
-        'max_roi_jitter_ratio', 'roi_jitter_mode', 'box_consistency_loss_weight'
+        'max_roi_jitter_ratio', 'roi_jitter_mode',
+        'box_consistency_loss_weight',
     ])):
   """Class holding the DeepMAC network configutration."""
 
@@ -125,6 +126,9 @@ def _get_deepmac_network_by_type(name, num_init_channels, mask_size=None):
       raise ValueError('Mask size must be set.')
     return FullyConnectedMaskHead(num_init_channels, mask_size)
 
+  elif name == 'embedding_distance_probability':
+    return tf.keras.layers.Lambda(lambda x: x)
+
   elif name.startswith('resnet'):
     return ResNetMaskNetwork(name, num_init_channels)
 
@@ -262,6 +266,25 @@ def fill_boxes(boxes, height, width):
   return tf.cast(filled_boxes, tf.float32)
 
 
+def embedding_distance_to_probability(x, y):
+  """Compute probability based on pixel-wise embedding distance.
+
+  Args:
+    x: [num_instances, height, width, dimension] float tensor input.
+    y: [num_instances, height, width, dimension] or
+      [num_instances, 1, 1, dimension] float tensor input. When the height
+      and width dimensions are 1, TF will broadcast it.
+
+  Returns:
+    dist: [num_instances, height, width, 1] A float tensor returning
+      the per-pixel probability. Pixels whose embeddings are close in
+      euclidean distance get a probability of close to 1.
+  """
+  diff = x - y
+  squared_dist = tf.reduce_sum(diff * diff, axis=3, keepdims=True)
+  return tf.exp(-squared_dist)
+
+
 class ResNetMaskNetwork(tf.keras.layers.Layer):
   """A small wrapper around ResNet blocks to predict masks."""
 
@@ -366,6 +389,16 @@ class MaskHeadNetwork(tf.keras.layers.Layer):
         network_type, num_init_channels, mask_size)
     self._use_instance_embedding = use_instance_embedding
 
+    self._network_type = network_type
+
+    if (self._use_instance_embedding and
+        (self._network_type == 'embedding_distance_probability')):
+      raise ValueError(('Cannot feed instance embedding to mask head when '
+                        'computing distance from instance embedding.'))
+
+    if network_type == 'embedding_distance_probability':
+      self.project_out = tf.keras.layers.Lambda(lambda x: x)
+    else:
       self.project_out = tf.keras.layers.Conv2D(
           filters=1, kernel_size=1, activation=None)
 
@@ -388,10 +421,9 @@ class MaskHeadNetwork(tf.keras.layers.Layer):
     height = tf.shape(pixel_embedding)[1]
     width = tf.shape(pixel_embedding)[2]
 
+    if self._use_instance_embedding:
       instance_embedding = instance_embedding[:, tf.newaxis, tf.newaxis, :]
       instance_embedding = tf.tile(instance_embedding, [1, height, width, 1])
-
-    if self._use_instance_embedding:
       inputs = tf.concat([pixel_embedding, instance_embedding], axis=3)
     else:
       inputs = pixel_embedding
@@ -400,6 +432,10 @@ class MaskHeadNetwork(tf.keras.layers.Layer):
     if isinstance(out, list):
       out = out[-1]
 
+    if self._network_type == 'embedding_distance_probability':
+      instance_embedding = instance_embedding[:, tf.newaxis, tf.newaxis, :]
+      out = embedding_distance_to_probability(instance_embedding, out)
+
     if out.shape[-1] > 1:
       out = self.project_out(out)
 
@@ -466,6 +502,25 @@ class DeepMACMetaArch(center_net_meta_arch.CenterNetMetaArch):
     if self._deepmac_params.mask_num_subsamples > 0:
       raise ValueError('Subsampling masks is currently not supported.')
 
+    if self._deepmac_params.network_type == 'embedding_distance_probability':
+      if self._deepmac_params.use_xy:
+        raise ValueError(
+            'Cannot use x/y coordinates when using embedding distance.')
+
+      pixel_embedding_dim = self._deepmac_params.pixel_embedding_dim
+      dim = self._deepmac_params.dim
+      if dim != pixel_embedding_dim:
+        raise ValueError(
+            'When using embedding distance mask head, '
+            f'pixel_embedding_dim({pixel_embedding_dim}) '
+            f'must be same as dim({dim}).')
+
+      loss = self._deepmac_params.classification_loss
+      if ((not isinstance(loss, losses.WeightedDiceClassificationLoss))
+          or (not loss.is_prediction_probability)):
+        raise ValueError('Only dice loss with is_prediction_probability=true '
+                         'is supported with embedding distance mask head.')
+
     super(DeepMACMetaArch, self).__init__(
         is_training=is_training, add_summaries=add_summaries,
         num_classes=num_classes, feature_extractor=feature_extractor,
@@ -909,6 +964,9 @@ class DeepMACMetaArch(center_net_meta_arch.CenterNetMetaArch):
       mask_logits = crop_masks_within_boxes(
           mask_logits, boxes, self._deepmac_params.postprocess_crop_size)
 
+    if self._deepmac_params.network_type == 'embedding_distance_probability':
+      masks_prob = mask_logits
+    else:
       masks_prob = tf.nn.sigmoid(mask_logits)
 
     return masks_prob

research/object_detection/meta_architectures/deepmac_meta_arch_test.py

@@ -61,7 +61,10 @@ class MockMaskNet(tf.keras.layers.Layer):
 
 
 def build_meta_arch(predict_full_resolution_masks=False, use_dice_loss=False,
-                    mask_num_subsamples=-1):
+                    use_instance_embedding=True, mask_num_subsamples=-1,
+                    network_type='hourglass10', use_xy=True,
+                    pixel_embedding_dim=2,
+                    dice_loss_prediction_probability=False):
   """Builds the DeepMAC meta architecture."""
 
   feature_extractor = DummyFeatureExtractor(
@@ -84,7 +87,9 @@ def build_meta_arch(predict_full_resolution_masks=False, use_dice_loss=False,
       use_labeled_classes=False)
 
   if use_dice_loss:
-    classification_loss = losses.WeightedDiceClassificationLoss(False)
+    classification_loss = losses.WeightedDiceClassificationLoss(
+        squared_normalization=False,
+        is_prediction_probability=dice_loss_prediction_probability)
   else:
     classification_loss = losses.WeightedSigmoidClassificationLoss()
 
@@ -92,13 +97,13 @@ def build_meta_arch(predict_full_resolution_masks=False, use_dice_loss=False,
       classification_loss=classification_loss,
       dim=8,
       task_loss_weight=1.0,
-      pixel_embedding_dim=2,
+      pixel_embedding_dim=pixel_embedding_dim,
       allowed_masked_classes_ids=[],
       mask_size=16,
       mask_num_subsamples=mask_num_subsamples,
-      use_xy=True,
-      network_type='hourglass10',
-      use_instance_embedding=True,
+      use_xy=use_xy,
+      network_type=network_type,
+      use_instance_embedding=use_instance_embedding,
       num_init_channels=8,
       predict_full_resolution_masks=predict_full_resolution_masks,
       postprocess_crop_size=128,
@@ -125,7 +130,7 @@ def build_meta_arch(predict_full_resolution_masks=False, use_dice_loss=False,
 
 
 @unittest.skipIf(tf_version.is_tf1(), 'Skipping TF2.X only test.')
-class DeepMACUtilsTest(tf.test.TestCase):
+class DeepMACUtilsTest(tf.test.TestCase, parameterized.TestCase):
 
   def test_subsample_trivial(self):
     """Test subsampling masks."""
@@ -169,12 +174,22 @@ class DeepMACUtilsTest(tf.test.TestCase):
         features, boxes, 32)
     self.assertEqual(output.shape, (5, 32, 32, 7))
 
+  def test_embedding_distance_prob_shape(self):
+    dist = deepmac_meta_arch.embedding_distance_to_probability(
+        tf.ones((4, 32, 32, 8)), tf.zeros((4, 32, 32, 8)))
+    self.assertEqual(dist.shape, (4, 32, 32, 1))
 
-@unittest.skipIf(tf_version.is_tf1(), 'Skipping TF2.X only test.')
-class DeepMACMetaArchTest(tf.test.TestCase):
+  @parameterized.parameters([1e-20, 1e20])
+  def test_embedding_distance_prob_value(self, value):
+    dist = deepmac_meta_arch.embedding_distance_to_probability(
+        tf.zeros((1, 1, 1, 8)), value + tf.zeros((1, 1, 1, 8))).numpy()
+    max_float = np.finfo(dist.dtype).max
+    self.assertLess(dist.max(), max_float)
+    self.assertGreater(dist.max(), -max_float)
 
-  def setUp(self):  # pylint:disable=g-missing-super-call
-    self.model = build_meta_arch()
+
+@unittest.skipIf(tf_version.is_tf1(), 'Skipping TF2.X only test.')
+class DeepMACMaskHeadTest(tf.test.TestCase):
 
   def test_mask_network(self):
     net = deepmac_meta_arch.MaskHeadNetwork('hourglass10', 8)
@@ -203,6 +218,38 @@ class DeepMACMetaArchTest(tf.test.TestCase):
     out = call_func(tf.zeros((2, 4)), tf.zeros((2, 32, 32, 16)), training=True)
     self.assertEqual(out.shape, (2, 32, 32))
 
+  def test_mask_network_embedding_distance_zero_dist(self):
+
+    net = deepmac_meta_arch.MaskHeadNetwork(
+        'embedding_distance_probability', num_init_channels=8,
+        use_instance_embedding=False)
+    call_func = tf.function(net.__call__)
+
+    out = call_func(tf.zeros((2, 7)), tf.zeros((2, 32, 32, 7)), training=True)
+    self.assertEqual(out.shape, (2, 32, 32))
+    self.assertAllGreater(out.numpy(), -np.inf)
+    self.assertAllLess(out.numpy(), np.inf)
+
+  def test_mask_network_embedding_distance_small_dist(self):
+
+    net = deepmac_meta_arch.MaskHeadNetwork(
+        'embedding_distance_probability', num_init_channels=-1,
+        use_instance_embedding=False)
+    call_func = tf.function(net.__call__)
+
+    out = call_func(1e6 + tf.zeros((2, 7)),
+                    tf.zeros((2, 32, 32, 7)), training=True)
+    self.assertEqual(out.shape, (2, 32, 32))
+    self.assertAllGreater(out.numpy(), -np.inf)
+    self.assertAllLess(out.numpy(), np.inf)
+
+
+@unittest.skipIf(tf_version.is_tf1(), 'Skipping TF2.X only test.')
+class DeepMACMetaArchTest(tf.test.TestCase, parameterized.TestCase):
+
+  def setUp(self):  # pylint:disable=g-missing-super-call
+    self.model = build_meta_arch()
+
   def test_get_mask_head_input(self):
 
     boxes = tf.constant([[0., 0., 0.25, 0.25], [0.75, 0.75, 1.0, 1.0]],
@@ -349,6 +396,37 @@ class DeepMACMetaArchTest(tf.test.TestCase):
     prob = tf.nn.sigmoid(0.9).numpy()
     self.assertAllClose(masks, prob * np.ones((2, 3, 16, 16)))
 
+  def test_postprocess_emb_dist(self):
+
+    model = build_meta_arch(network_type='embedding_distance_probability',
+                            use_instance_embedding=False,
+                            use_xy=False, pixel_embedding_dim=8,
+                            use_dice_loss=True,
+                            dice_loss_prediction_probability=True)
+    boxes = np.zeros((2, 3, 4), dtype=np.float32)
+    boxes[:, :, [0, 2]] = 0.0
+    boxes[:, :, [1, 3]] = 8.0
+    boxes = tf.constant(boxes)
+
+    masks = model._postprocess_masks(
+        boxes, tf.zeros((2, 32, 32, 2)), tf.zeros((2, 32, 32, 2)))
+    self.assertEqual(masks.shape, (2, 3, 16, 16))
+
+  def test_postprocess_emb_dist_fullres(self):
+
+    model = build_meta_arch(network_type='embedding_distance_probability',
+                            predict_full_resolution_masks=True,
+                            use_instance_embedding=False,
+                            pixel_embedding_dim=8, use_xy=False,
+                            use_dice_loss=True,
+                            dice_loss_prediction_probability=True)
+    boxes = np.zeros((2, 3, 4), dtype=np.float32)
+    boxes = tf.constant(boxes)
+
+    masks = model._postprocess_masks(
+        boxes, tf.zeros((2, 32, 32, 2)), tf.zeros((2, 32, 32, 2)))
+    self.assertEqual(masks.shape, (2, 3, 128, 128))
+
   def test_postprocess_no_crop_resize_shape(self):
 
     model = build_meta_arch(predict_full_resolution_masks=True)
@@ -494,7 +572,7 @@ class FullyConnectedMaskHeadTest(tf.test.TestCase):
 class ResNetMaskHeadTest(tf.test.TestCase, parameterized.TestCase):
 
   @parameterized.parameters(['resnet4', 'resnet8', 'resnet20'])
-  def test_pass(self, name):
+  def test_forward(self, name):
     net = deepmac_meta_arch.ResNetMaskNetwork(name, 8)
     out = net(tf.zeros((3, 32, 32, 16)))
     self.assertEqual(out.shape[:3], (3, 32, 32))

research/object_detection/protos/losses.proto

@@ -231,6 +231,10 @@ message WeightedDiceClassificationLoss {
   // If set, we square the probabilities in the denominator term used for
   // normalization.
   optional bool squared_normalization = 1 [default=false];
+
+  // Whether or not the input prediction to the loss function is a
+  // probability. If not, the input is to be interpreted as logit
+  optional bool is_prediction_probability = 2 [default=false];
 }