Minor change to mask target assigner to optionally scale boxes before assigning dense loss weights.

PiperOrigin-RevId: 382773317

research/object_detection/core/target_assigner.py

@@ -2003,8 +2003,20 @@ def _resize_masks(masks, height, width, method):
 class CenterNetMaskTargetAssigner(object):
   """Wrapper to compute targets for segmentation masks."""
 
-  def __init__(self, stride):
+  def __init__(self, stride, boxes_scale=1.0):
+    """Constructor.
+
+    Args:
+      stride: The stride of the network. Targets are assigned at the output
+        stride.
+      boxes_scale: Scale to apply to boxes before producing mask weights. This
+        is meant to ensure the full object region is properly weighted prior to
+        applying loss. A value of ~1.05 is typically applied when object regions
+        should be blacked out (perhaps because valid groundtruth masks are not
+        present).
+    """
     self._stride = stride
+    self._boxes_scale = boxes_scale
 
   def assign_segmentation_targets(
       self, gt_masks_list, gt_classes_list, gt_boxes_list=None,
@@ -2072,7 +2084,7 @@ class CenterNetMaskTargetAssigner(object):
         segmentation_weight_for_image = (
             ta_utils.blackout_pixel_weights_by_box_regions(
                 output_height, output_width, boxes_absolute.get(), blackout,
-                weights=gt_mask_weights))
+                weights=gt_mask_weights, boxes_scale=self._boxes_scale))
         segmentation_weights_list.append(segmentation_weight_for_image)
       else:
         segmentation_weights_list.append(tf.ones((output_height, output_width),

research/object_detection/meta_architectures/center_net_meta_arch.py

@@ -2655,7 +2655,7 @@ class CenterNetMetaArch(model.DetectionModel):
                 per_keypoint_depth=kp_params.per_keypoint_depth))
     if self._mask_params is not None:
       target_assigners[SEGMENTATION_TASK] = (
-          cn_assigner.CenterNetMaskTargetAssigner(stride))
+          cn_assigner.CenterNetMaskTargetAssigner(stride, boxes_scale=1.05))
     if self._densepose_params is not None:
       dp_stride = 1 if self._densepose_params.upsample_to_input_res else stride
       target_assigners[DENSEPOSE_TASK] = (

research/object_detection/utils/target_assigner_utils.py

@@ -16,6 +16,8 @@
 
 import tensorflow.compat.v1 as tf
 
+from object_detection.core import box_list
+from object_detection.core import box_list_ops
 from object_detection.utils import shape_utils
 
 
@@ -290,7 +292,8 @@ def get_valid_keypoint_mask_for_class(keypoint_coordinates,
 
 
 def blackout_pixel_weights_by_box_regions(height, width, boxes, blackout,
-                                          weights=None):
+                                          weights=None,
+                                          boxes_scale=1.0):
   """Apply weights at pixel locations.
 
   This function is used to generate the pixel weight mask (usually in the output
@@ -332,6 +335,10 @@ def blackout_pixel_weights_by_box_regions(height, width, boxes, blackout,
       a value to apply in each box region. Note that if blackout=True for a
       given box, the weight will be zero. If None, all weights are assumed to be
       1.
+    boxes_scale: The amount to scale the height/width of the boxes before
+      constructing the blackout regions. This is often useful to guarantee that
+      the proper weight fully covers the object boxes/masks during supervision,
+      as shifting might occur during image resizing, network stride, etc.
 
   Returns:
     A float tensor with shape [height, width] where all values within the
@@ -347,6 +354,10 @@ def blackout_pixel_weights_by_box_regions(height, width, boxes, blackout,
   (y_grid, x_grid) = image_shape_to_grids(height, width)
   y_grid = tf.expand_dims(y_grid, axis=0)
   x_grid = tf.expand_dims(x_grid, axis=0)
+  boxlist = box_list.BoxList(boxes)
+  boxlist = box_list_ops.scale_height_width(
+      boxlist, y_scale=boxes_scale, x_scale=boxes_scale)
+  boxes = boxlist.get()
   y_min = tf.expand_dims(boxes[:, 0:1], axis=-1)
   x_min = tf.expand_dims(boxes[:, 1:2], axis=-1)
   y_max = tf.expand_dims(boxes[:, 2:3], axis=-1)