Use functions from util map.

PiperOrigin-RevId: 325621239

research/object_detection/meta_architectures/center_net_meta_arch.py

@@ -2583,6 +2583,9 @@ class CenterNetMetaArch(model.DetectionModel):
       detections: a dictionary containing the following fields
         detection_boxes - A tensor of shape [batch, max_detections, 4]
           holding the predicted boxes.
+        detection_boxes_strided: A tensor of shape [batch_size, num_detections,
+          4] holding the predicted boxes in absolute coordinates of the
+          feature extractor's final layer output.
         detection_scores: A tensor of shape [batch, max_detections] holding
           the predicted score for each box.
         detection_classes: An integer tensor of shape [batch, max_detections]
@@ -2626,6 +2629,7 @@ class CenterNetMetaArch(model.DetectionModel):
         fields.DetectionResultFields.detection_scores: scores,
         fields.DetectionResultFields.detection_classes: classes,
         fields.DetectionResultFields.num_detections: num_detections,
+        'detection_boxes_strided': boxes_strided
     }
 
     if self._kp_params_dict:

research/object_detection/model_lib_v2.py

@@ -28,7 +28,6 @@ import tensorflow.compat.v2 as tf2
 from object_detection import eval_util
 from object_detection import inputs
 from object_detection import model_lib
-from object_detection.builders import model_builder
 from object_detection.builders import optimizer_builder
 from object_detection.core import standard_fields as fields
 from object_detection.protos import train_pb2
@@ -503,7 +502,7 @@ def train_loop(
   # Build the model, optimizer, and training input
   strategy = tf.compat.v2.distribute.get_strategy()
   with strategy.scope():
-    detection_model = model_builder.build(
+    detection_model = MODEL_BUILD_UTIL_MAP['detection_model_fn_base'](
         model_config=model_config, is_training=True)
 
     def train_dataset_fn(input_context):
@@ -939,7 +938,7 @@ def eval_continuously(
   if kwargs['use_bfloat16']:
     tf.compat.v2.keras.mixed_precision.experimental.set_policy('mixed_bfloat16')
 
-  detection_model = model_builder.build(
+  detection_model = MODEL_BUILD_UTIL_MAP['detection_model_fn_base'](
       model_config=model_config, is_training=True)
 
   # Create the inputs.

research/object_detection/protos/center_net.proto

@@ -244,6 +244,42 @@ message CenterNet {
     optional ClassificationLoss classification_loss = 5;
   }
   optional TrackEstimation track_estimation_task = 10;
+
+  // BEGIN GOOGLE-INTERNAL
+  // Experimental Occupancy network head, use with caution.
+  message OccupancyNetMaskPrediction {
+    // The loss used for penalizing mask predictions.
+    optional ClassificationLoss classification_loss = 1;
+
+    // Number of points to sample within a box while training occupancy net.
+    optional int32 num_samples = 2 [default = 1000];
+
+    // The dimension of the occupancy embedding.
+    optional int32 dim = 3 [default = 256];
+
+    // Weight of occupancy embedding loss.
+    optional float task_loss_weight = 4 [default = 1.0];
+
+    // The stride in pixels at test time when computing the mask. THis is
+    // useful is computing the full mask is too expensive.
+    optional int32 mask_stride = 5 [default = 1];
+
+    // If set, concatenate the occupancy embedding features to (x, y)
+    // coordinates before feeding it to the occupancy network head.
+    optional bool concat_features = 6 [default = true];
+
+    // If set to a positive value, defines the length to which the embedding
+    // is clipped before concatenating to the (x, y) coordinates when
+    // concat_features=true.
+    optional int32 concat_clip = 7 [default = -1];
+
+    // The probability threshold to apply for masks to output a binary mask.
+    optional float mask_prob_threshold = 8 [default = 0.5];
+  }
+
+  optional OccupancyNetMaskPrediction occupancy_net_mask_prediction = 11;
+  // EBD GOOGLE-INTERNAL
+
 }
 
 message CenterNetFeatureExtractor {