compute small instance weights based on mask area

official/vision/beta/projects/panoptic_maskrcnn/configs/panoptic_deeplab.py

@@ -45,6 +45,8 @@ class Parser(hyperparams.Config):
   aug_scale_max: float = 1.0
   aug_rand_hflip: bool = True
   sigma: float = 8.0
+  small_instance_area_threshold: int = 4096
+  small_instance_weight: float = 3.0
   dtype = 'float32'
 
 @dataclasses.dataclass
@@ -92,8 +94,8 @@ class PanopticDeeplabPostProcessor(hyperparams.Config):
   label_divisor: int = 256 * 256 * 256
   stuff_area_limit: int = 4096
   ignore_label: int = 0
-  nms_kernel: int = 41
-  keep_k_centers: int = 400
+  nms_kernel: int = 7
+  keep_k_centers: int = 200
   rescale_predictions: bool = True
 
 @dataclasses.dataclass
@@ -119,7 +121,6 @@ class Losses(hyperparams.Config):
   ignore_label: int = 0
   class_weights: List[float] = dataclasses.field(default_factory=list)
   l2_weight_decay: float = 1e-4
-  use_groundtruth_dimension: bool = True
   top_k_percent_pixels: float = 0.15
   segmentation_loss_weight: float = 1.0
   center_heatmap_loss_weight: float = 200
@@ -235,7 +236,6 @@ def panoptic_deeplab_coco() -> cfg.ExperimentConfig:
               label_smoothing=0.0,
               ignore_label=ignore_label,
               l2_weight_decay=0.0,
-              use_groundtruth_dimension=True,
               top_k_percent_pixels=0.2,
               segmentation_loss_weight=1.0,
               center_heatmap_loss_weight=200,
@@ -248,7 +248,9 @@ def panoptic_deeplab_coco() -> cfg.ExperimentConfig:
                   aug_scale_min=0.5,
                   aug_scale_max=1.5,
                   aug_rand_hflip=True,
-                  sigma=8.0)),
+                  sigma=8.0,
+                  small_instance_area_threshold=4096,
+                  small_instance_weight=3.0)),
           validation_data=DataConfig(
               input_path=os.path.join(_COCO_INPUT_PATH_BASE, 'val*'),
               is_training=False,
@@ -259,15 +261,17 @@ def panoptic_deeplab_coco() -> cfg.ExperimentConfig:
                   aug_scale_min=1.0,
                   aug_scale_max=1.0,
                   aug_rand_hflip=False,
-                  sigma=8.0),
+                  sigma=8.0,
+                  small_instance_area_threshold=4096,
+                  small_instance_weight=3.0),
               drop_remainder=False),
           evaluation=Evaluation(
               ignored_label=ignore_label,
               max_instances_per_category=256,
-              offset=256 * 256 * 256,
+              offset=256*256*256,
               is_thing=is_thing,
               rescale_predictions=True,
-              report_per_class_pq=True,
+              report_per_class_pq=False,
               report_per_class_iou=False,
               report_train_mean_iou=False)),
       trainer=cfg.TrainerConfig(

official/vision/beta/projects/panoptic_maskrcnn/dataloaders/panoptic_deeplab_input.py

@@ -83,6 +83,8 @@ class Parser(parser.Parser):
       aug_scale_min=1.0,
       aug_scale_max=1.0,
       sigma=8.0,
+      small_instance_area_threshold=4096,
+      small_instance_weight=3.0,
       dtype='float32'):
     """Initializes parameters for parsing annotations in the dataset.
 
@@ -104,6 +106,8 @@ class Parser(parser.Parser):
         data augmentation during training.
       sigma: `float`, standard deviation for generating 2D Gaussian to encode
         centers.
+      small_instance_area_threshold: `int`, 
+      small_instance_weight: `float`, 
       dtype: `str`, data type. One of {`bfloat16`, `float32`, `float16`}.
     """
     self._output_size = output_size
@@ -127,6 +131,9 @@ class Parser(parser.Parser):
     self._gaussian, self._gaussian_size = _compute_gaussian_from_std(
         self._sigma)
     self._gaussian = tf.reshape(self._gaussian, shape=[-1])
+    self._small_instance_area_threshold = small_instance_area_threshold
+    self._small_instance_weight = small_instance_weight
+
 
   def _resize_and_crop_mask(self, mask, image_info, is_training):
     """Resizes and crops mask using `image_info` dict"""
@@ -194,7 +201,9 @@ class Parser(parser.Parser):
         image_info,
         is_training=is_training)
 
-    instance_centers_heatmap, instance_centers_offset = self._encode_centers_and_offets(
+    (instance_centers_heatmap,
+     instance_centers_offset,
+     semantic_weights) = self._encode_centers_and_offets(
          instance_mask=instance_mask[:, :, 0])
 
     # Cast image and labels as self._dtype
@@ -216,6 +225,7 @@ class Parser(parser.Parser):
         'instance_mask': instance_mask,
         'instance_centers_heatmap': instance_centers_heatmap,
         'instance_centers_offset': instance_centers_offset,
+        'semantic_weights': semantic_weights,
         'valid_mask': valid_mask,
         'things_mask': things_mask,
         'image_info': image_info
@@ -259,6 +269,9 @@ class Parser(parser.Parser):
     centers_offset_x = tf.zeros(
         shape=[height, width],
         dtype=tf.float32)
+    semantic_weights = tf.ones(
+        shape=[height, width],
+        dtype=tf.float32)
 
     unique_instance_ids, _ = tf.unique(tf.reshape(instance_mask, [-1]))
 
@@ -270,11 +283,18 @@ class Parser(parser.Parser):
         continue
 
       mask = tf.equal(instance_mask, instance_id)
+      mask_area = tf.reduce_sum(tf.cast(mask, dtype=tf.float32))
       mask_indices = tf.cast(tf.where(mask), dtype=tf.float32)
       mask_center = tf.reduce_mean(mask_indices, axis=0)
       mask_center_y = tf.cast(tf.round(mask_center[0]), dtype=tf.int32)
       mask_center_x = tf.cast(tf.round(mask_center[1]), dtype=tf.int32)
 
+      if mask_area < self._small_instance_area_threshold:
+        semantic_weights = tf.where(
+            mask,
+            self._small_instance_weight,
+            semantic_weights)
+
       gaussian_size = self._gaussian_size
       indices_y = tf.range(mask_center_y, mask_center_y + gaussian_size)
       indices_x = tf.range(mask_center_x, mask_center_x + gaussian_size)
@@ -308,4 +328,6 @@ class Parser(parser.Parser):
         [centers_offset_y, centers_offset_x], 
         axis=-1)
     
-    return instance_centers_heatmap, instance_centers_offset
+    return (instance_centers_heatmap,
+            instance_centers_offset,
+            semantic_weights)