Plumb `is_crowd` flag into COCO Mask eval

PiperOrigin-RevId: 337134017

research/object_detection/metrics/coco_evaluation.py

@@ -997,12 +997,27 @@ class CocoMaskEvaluator(object_detection_evaluation.DetectionEvaluator):
           [num_boxes, image_height, image_width] containing groundtruth masks
           corresponding to the boxes. The elements of the array must be in
           {0, 1}.
+        InputDataFields.groundtruth_is_crowd (optional): integer numpy array of
+          shape [num_boxes] containing iscrowd flag for groundtruth boxes.
+        InputDataFields.groundtruth_area (optional): float numpy array of
+          shape [num_boxes] containing the area (in the original absolute
+          coordinates) of the annotated object.
     """
     if image_id in self._image_id_to_mask_shape_map:
       tf.logging.warning('Ignoring ground truth with image id %s since it was '
                          'previously added', image_id)
       return
 
+    # Drop optional fields if empty tensor.
+    groundtruth_is_crowd = groundtruth_dict.get(
+        standard_fields.InputDataFields.groundtruth_is_crowd)
+    groundtruth_area = groundtruth_dict.get(
+        standard_fields.InputDataFields.groundtruth_area)
+    if groundtruth_is_crowd is not None and not groundtruth_is_crowd.shape[0]:
+      groundtruth_is_crowd = None
+    if groundtruth_area is not None and not groundtruth_area.shape[0]:
+      groundtruth_area = None
+
     groundtruth_instance_masks = groundtruth_dict[
         standard_fields.InputDataFields.groundtruth_instance_masks]
     groundtruth_instance_masks = convert_masks_to_binary(
@@ -1018,7 +1033,9 @@ class CocoMaskEvaluator(object_detection_evaluation.DetectionEvaluator):
             groundtruth_classes=groundtruth_dict[standard_fields.
                                                  InputDataFields.
                                                  groundtruth_classes],
-            groundtruth_masks=groundtruth_instance_masks))
+            groundtruth_masks=groundtruth_instance_masks,
+            groundtruth_is_crowd=groundtruth_is_crowd,
+            groundtruth_area=groundtruth_area))
     self._annotation_id += groundtruth_dict[standard_fields.InputDataFields.
                                             groundtruth_boxes].shape[0]
     self._image_id_to_mask_shape_map[image_id] = groundtruth_dict[

research/object_detection/metrics/coco_evaluation_test.py

@@ -1556,6 +1556,41 @@ class CocoMaskEvaluationTest(tf.test.TestCase):
     self.assertFalse(coco_evaluator._groundtruth_list)
     self.assertFalse(coco_evaluator._detection_masks_list)
 
+  def testGetOneMAPWithMatchingGroundtruthAndDetectionsSkipCrowd(self):
+    """Tests computing mAP with is_crowd GT boxes skipped."""
+    coco_evaluator = coco_evaluation.CocoMaskEvaluator(
+        _get_categories_list())
+    coco_evaluator.add_single_ground_truth_image_info(
+        image_id='image1',
+        groundtruth_dict={
+            standard_fields.InputDataFields.groundtruth_boxes:
+                np.array([[100., 100., 200., 200.], [99., 99., 200., 200.]]),
+            standard_fields.InputDataFields.groundtruth_classes:
+                np.array([1, 2]),
+            standard_fields.InputDataFields.groundtruth_is_crowd:
+                np.array([0, 1]),
+            standard_fields.InputDataFields.groundtruth_instance_masks:
+                np.concatenate(
+                    [np.pad(np.ones([1, 100, 100], dtype=np.uint8),
+                            ((0, 0), (100, 56), (100, 56)), mode='constant'),
+                     np.pad(np.ones([1, 101, 101], dtype=np.uint8),
+                            ((0, 0), (99, 56), (99, 56)), mode='constant')],
+                    axis=0)
+        })
+    coco_evaluator.add_single_detected_image_info(
+        image_id='image1',
+        detections_dict={
+            standard_fields.DetectionResultFields.detection_scores:
+                np.array([.8]),
+            standard_fields.DetectionResultFields.detection_classes:
+                np.array([1]),
+            standard_fields.DetectionResultFields.detection_masks:
+                np.pad(np.ones([1, 100, 100], dtype=np.uint8),
+                       ((0, 0), (100, 56), (100, 56)), mode='constant')
+        })
+    metrics = coco_evaluator.evaluate()
+    self.assertAlmostEqual(metrics['DetectionMasks_Precision/mAP'], 1.0)
+
 
 @unittest.skipIf(tf_version.is_tf2(), 'Only Supported in TF1.X')
 class CocoMaskEvaluationPyFuncTest(tf.test.TestCase):