Add multilevel functions dealing with multilevel features.

research/object_detection/meta_architectures/faster_rcnn_meta_arch.py

@@ -1963,7 +1963,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
     second stage box classifier for each proposal.
 
     Args:
-      features_to_crop: A float32 tensor with shape
+      features_to_crop: A list of float32 tensor with shape
         [batch_size, height, width, depth]
       proposal_boxes_normalized: A float32 tensor with shape [batch_size,
         num_proposals, box_code_size] containing proposal boxes in
@@ -1973,9 +1973,15 @@ class FasterRCNNMetaArch(model.DetectionModel):
     Returns:
       A float32 tensor with shape [K, new_height, new_width, depth].
     """
+    num_levels = len(features_to_crop)
+    box_levels = None
+    if num_levels != 1:
+      # If there are mutiple levels to select, get the box levels 
+      box_levels = ops.fpn_feature_levels(num_levels, num_levels - 1,
+                                          1.0/224, proposal_boxes_normalized)
     cropped_regions = self._flatten_first_two_dimensions(
         self._crop_and_resize_fn(
-            features_to_crop, proposal_boxes_normalized,
+            features_to_crop, proposal_boxes_normalized, box_levels,
             [self._initial_crop_size, self._initial_crop_size]))
     return self._maxpool_layer(cropped_regions)
 

research/object_detection/utils/spatial_transform_ops.py

@@ -411,6 +411,19 @@ def multilevel_roi_align(features, boxes, box_levels, output_size,
     return features_per_box
 
 
+def multilevel_native_crop_and_resize(images, boxes, box_levels, 
+                                      crop_size, scope=None):
+  #FIXME: fix docstring
+  """doc string."""
+  if not box_levels:
+    return native_crop_and_resize(images[0], boxes, crop_size, scope=None)
+  croped_feature_list = []
+  for level, image in enumerate(images):
+    level_boxes = tf.gather(boxes, box_levels == (level-1))
+    cropped = native_crop_and_resize(image, level_boxes, crop_size)
+    croped_feature_list.append(cropped)
+  return tf.concat(croped_feature_list, axis=0)
+
 def native_crop_and_resize(image, boxes, crop_size, scope=None):
   """Same as `matmul_crop_and_resize` but uses tf.image.crop_and_resize."""
   def get_box_inds(proposals):
@@ -430,6 +443,20 @@ def native_crop_and_resize(image, boxes, crop_size, scope=None):
                              tf.shape(cropped_regions)[1:]], axis=0)
     return tf.reshape(cropped_regions, final_shape)
 
+def multilevel_matmul_crop_and_resize(images, boxes, box_levels, crop_size, 
+                                      extrapolation_value=0.0, scope=None):
+  #FIXME: fix docstring
+  """doc string."""
+  with tf.name_scope(scope, 'MatMulCropAndResize'):
+    if box_levels is None:
+      box_levels = tf.zeros(tf.shape(boxes)[:2], dtype=tf.int32)
+    return multilevel_roi_align([image],
+                                boxes,
+                                box_levels,
+                                crop_size,
+                                align_corners=True,
+                                extrapolation_value=extrapolation_value)
+
 
 def matmul_crop_and_resize(image, boxes, crop_size, extrapolation_value=0.0,
                            scope=None):