Pass in is_crowd for coco evaluation.

PiperOrigin-RevId: 187527188

research/object_detection/metrics/coco_evaluation.py

@@ -74,23 +74,30 @@ class CocoDetectionEvaluator(object_detection_evaluation.DetectionEvaluator):
           [ymin, xmin, ymax, xmax] in absolute image coordinates.
         InputDataFields.groundtruth_classes: integer numpy array of shape
           [num_boxes] containing 1-indexed groundtruth classes for the boxes.
+        InputDataFields.groundtruth_is_crowd (optional): integer numpy array of
+          shape [num_boxes] containing iscrowd flag for groundtruth boxes.
     """
     if image_id in self._image_ids:
       tf.logging.warning('Ignoring ground truth with image id %s since it was '
                          'previously added', image_id)
       return
 
+    groundtruth_is_crowd = groundtruth_dict.get(
+        standard_fields.InputDataFields.groundtruth_is_crowd)
+    # Drop groundtruth_is_crowd if empty tensor.
+    if groundtruth_is_crowd is not None and not groundtruth_is_crowd.shape[0]:
+      groundtruth_is_crowd = None
+
     self._groundtruth_list.extend(
-        coco_tools.
-        ExportSingleImageGroundtruthToCoco(
+        coco_tools.ExportSingleImageGroundtruthToCoco(
             image_id=image_id,
             next_annotation_id=self._annotation_id,
             category_id_set=self._category_id_set,
-            groundtruth_boxes=groundtruth_dict[standard_fields.InputDataFields.
-                                               groundtruth_boxes],
-            groundtruth_classes=groundtruth_dict[standard_fields.
-                                                 InputDataFields.
-                                                 groundtruth_classes]))
+            groundtruth_boxes=groundtruth_dict[
+                standard_fields.InputDataFields.groundtruth_boxes],
+            groundtruth_classes=groundtruth_dict[
+                standard_fields.InputDataFields.groundtruth_classes],
+            groundtruth_is_crowd=groundtruth_is_crowd))
     self._annotation_id += groundtruth_dict[standard_fields.InputDataFields.
                                             groundtruth_boxes].shape[0]
     self._image_ids[image_id] = False

research/object_detection/metrics/coco_evaluation_test.py

@@ -86,6 +86,78 @@ class CocoDetectionEvaluationTest(tf.test.TestCase):
     metrics = coco_evaluator.evaluate()
     self.assertAlmostEqual(metrics['DetectionBoxes_Precision/mAP'], 1.0)
 
+  def testGetOneMAPWithMatchingGroundtruthAndDetectionsSkipCrowd(self):
+    """Tests computing mAP with is_crowd GT boxes skipped."""
+    category_list = [{
+        'id': 0,
+        'name': 'person'
+    }, {
+        'id': 1,
+        'name': 'cat'
+    }, {
+        'id': 2,
+        'name': 'dog'
+    }]
+    coco_evaluator = coco_evaluation.CocoDetectionEvaluator(category_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])
+        })
+    coco_evaluator.add_single_detected_image_info(
+        image_id='image1',
+        detections_dict={
+            standard_fields.DetectionResultFields.detection_boxes:
+                np.array([[100., 100., 200., 200.]]),
+            standard_fields.DetectionResultFields.detection_scores:
+                np.array([.8]),
+            standard_fields.DetectionResultFields.detection_classes:
+                np.array([1])
+        })
+    metrics = coco_evaluator.evaluate()
+    self.assertAlmostEqual(metrics['DetectionBoxes_Precision/mAP'], 1.0)
+
+  def testGetOneMAPWithMatchingGroundtruthAndDetectionsEmptyCrowd(self):
+    """Tests computing mAP with empty is_crowd array passed in."""
+    category_list = [{
+        'id': 0,
+        'name': 'person'
+    }, {
+        'id': 1,
+        'name': 'cat'
+    }, {
+        'id': 2,
+        'name': 'dog'
+    }]
+    coco_evaluator = coco_evaluation.CocoDetectionEvaluator(category_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.]]),
+            standard_fields.InputDataFields.groundtruth_classes:
+                np.array([1]),
+            standard_fields.InputDataFields.groundtruth_is_crowd:
+                np.array([])
+        })
+    coco_evaluator.add_single_detected_image_info(
+        image_id='image1',
+        detections_dict={
+            standard_fields.DetectionResultFields.detection_boxes:
+                np.array([[100., 100., 200., 200.]]),
+            standard_fields.DetectionResultFields.detection_scores:
+                np.array([.8]),
+            standard_fields.DetectionResultFields.detection_classes:
+                np.array([1])
+        })
+    metrics = coco_evaluator.evaluate()
+    self.assertAlmostEqual(metrics['DetectionBoxes_Precision/mAP'], 1.0)
+
   def testRejectionOnDuplicateGroundtruth(self):
     """Tests that groundtruth cannot be added more than once for an image."""
     categories = [{'id': 1, 'name': 'cat'},

research/object_detection/metrics/coco_tools.py

@@ -327,7 +327,8 @@ def ExportSingleImageGroundtruthToCoco(image_id,
                                        category_id_set,
                                        groundtruth_boxes,
                                        groundtruth_classes,
-                                       groundtruth_masks=None):
+                                       groundtruth_masks=None,
+                                       groundtruth_is_crowd=None):
   """Export groundtruth of a single image to COCO format.
 
   This function converts groundtruth detection annotations represented as numpy
@@ -338,8 +339,7 @@ def ExportSingleImageGroundtruthToCoco(image_id,
   groundtruth_classes[i] are associated with the same groundtruth annotation.
 
   In the exported result, "area" fields are always set to the area of the
-  groundtruth bounding box and "iscrowd" fields are always set to 0.
-  TODO(jonathanhuang): pass in "iscrowd" array for evaluating on COCO dataset.
+  groundtruth bounding box.
 
   Args:
     image_id: a unique image identifier either of type integer or string.
@@ -352,6 +352,8 @@ def ExportSingleImageGroundtruthToCoco(image_id,
     groundtruth_classes: numpy array (int) with shape [num_gt_boxes]
     groundtruth_masks: optional uint8 numpy array of shape [num_detections,
       image_height, image_width] containing detection_masks.
+    groundtruth_is_crowd: optional numpy array (int) with shape [num_gt_boxes]
+      indicating whether groundtruth boxes are crowd.
 
   Returns:
     a list of groundtruth annotations for a single image in the COCO format.
@@ -379,17 +381,27 @@ def ExportSingleImageGroundtruthToCoco(image_id,
                      'Classes shape: %d. Boxes shape: %d. Image ID: %s' % (
                          groundtruth_classes.shape[0],
                          groundtruth_boxes.shape[0], image_id))
+  has_is_crowd = groundtruth_is_crowd is not None
+  if has_is_crowd and len(groundtruth_is_crowd.shape) != 1:
+    raise ValueError('groundtruth_is_crowd is expected to be of rank 1.')
   groundtruth_list = []
   for i in range(num_boxes):
     if groundtruth_classes[i] in category_id_set:
+      iscrowd = groundtruth_is_crowd[i] if has_is_crowd else 0
       export_dict = {
-          'id': next_annotation_id + i,
-          'image_id': image_id,
-          'category_id': int(groundtruth_classes[i]),
-          'bbox': list(_ConvertBoxToCOCOFormat(groundtruth_boxes[i, :])),
-          'area': float((groundtruth_boxes[i, 2] - groundtruth_boxes[i, 0]) *
-                        (groundtruth_boxes[i, 3] - groundtruth_boxes[i, 1])),
-          'iscrowd': 0
+          'id':
+              next_annotation_id + i,
+          'image_id':
+              image_id,
+          'category_id':
+              int(groundtruth_classes[i]),
+          'bbox':
+              list(_ConvertBoxToCOCOFormat(groundtruth_boxes[i, :])),
+          'area':
+              float((groundtruth_boxes[i, 2] - groundtruth_boxes[i, 0]) *
+                    (groundtruth_boxes[i, 3] - groundtruth_boxes[i, 1])),
+          'iscrowd':
+              iscrowd
       }
       if groundtruth_masks is not None:
         export_dict['segmentation'] = _RleCompress(groundtruth_masks[i])

research/object_detection/metrics/coco_tools_test.py

@@ -248,7 +248,11 @@ class CocoToolsTest(tf.test.TestCase):
                            [0, 0, .5, .5],
                            [.5, .5, .5, .5]], dtype=np.float32)
     classes = np.array([1, 2, 3], dtype=np.int32)
+    is_crowd = np.array([0, 1, 0], dtype=np.int32)
     next_annotation_id = 1
+    expected_counts = ['04', '31', '4']
+
+    # Tests exporting without passing in is_crowd (for backward compatibility).
     coco_annotations = coco_tools.ExportSingleImageGroundtruthToCoco(
         image_id='first_image',
         category_id_set=set([1, 2, 3]),
@@ -256,7 +260,6 @@ class CocoToolsTest(tf.test.TestCase):
         groundtruth_boxes=boxes,
         groundtruth_classes=classes,
         groundtruth_masks=masks)
-    expected_counts = ['04', '31', '4']
     for i, annotation in enumerate(coco_annotations):
       self.assertEqual(annotation['segmentation']['counts'],
                        expected_counts[i])
@@ -267,6 +270,26 @@ class CocoToolsTest(tf.test.TestCase):
       self.assertEqual(annotation['category_id'], classes[i])
       self.assertEqual(annotation['id'], i + next_annotation_id)
 
+    # Tests exporting with is_crowd.
+    coco_annotations = coco_tools.ExportSingleImageGroundtruthToCoco(
+        image_id='first_image',
+        category_id_set=set([1, 2, 3]),
+        next_annotation_id=next_annotation_id,
+        groundtruth_boxes=boxes,
+        groundtruth_classes=classes,
+        groundtruth_masks=masks,
+        groundtruth_is_crowd=is_crowd)
+    for i, annotation in enumerate(coco_annotations):
+      self.assertEqual(annotation['segmentation']['counts'],
+                       expected_counts[i])
+      self.assertTrue(np.all(np.equal(mask.decode(
+          annotation['segmentation']), masks[i])))
+      self.assertTrue(np.all(np.isclose(annotation['bbox'], coco_boxes[i])))
+      self.assertEqual(annotation['image_id'], 'first_image')
+      self.assertEqual(annotation['category_id'], classes[i])
+      self.assertEqual(annotation['iscrowd'], is_crowd[i])
+      self.assertEqual(annotation['id'], i + next_annotation_id)
+
 
 if __name__ == '__main__':
   tf.test.main()