Making NMS flag visible through ObjectDetectionEvaluator constructor. Note that

by default NMS is off.

PiperOrigin-RevId: 347613472

research/object_detection/utils/object_detection_evaluation.py

@@ -159,7 +159,9 @@ class ObjectDetectionEvaluator(DetectionEvaluator):
                metric_prefix=None,
                use_weighted_mean_ap=False,
                evaluate_masks=False,
-               group_of_weight=0.0):
+               group_of_weight=0.0,
+               nms_iou_threshold=1.0,
+               nms_max_output_boxes=10000):
     """Constructor.
 
     Args:
@@ -187,6 +189,8 @@ class ObjectDetectionEvaluator(DetectionEvaluator):
         matching_iou_threshold, weight group_of_weight is added to true
         positives. Consequently, if no detection falls within a group-of box,
         weight group_of_weight is added to false negatives.
+      nms_iou_threshold: NMS IoU threashold.
+      nms_max_output_boxes: maximal number of boxes after NMS.
 
     Raises:
       ValueError: If the category ids are not 1-indexed.
@@ -202,6 +206,8 @@ class ObjectDetectionEvaluator(DetectionEvaluator):
     self._label_id_offset = 1
     self._evaluate_masks = evaluate_masks
     self._group_of_weight = group_of_weight
+    self._nms_iou_threshold = nms_iou_threshold
+    self._nms_max_output_boxes = nms_max_output_boxes
     self._evaluation = ObjectDetectionEvaluation(
         num_groundtruth_classes=self._num_classes,
         matching_iou_threshold=self._matching_iou_threshold,
@@ -209,7 +215,9 @@ class ObjectDetectionEvaluator(DetectionEvaluator):
         recall_upper_bound=self._recall_upper_bound,
         use_weighted_mean_ap=self._use_weighted_mean_ap,
         label_id_offset=self._label_id_offset,
-        group_of_weight=self._group_of_weight)
+        group_of_weight=self._group_of_weight,
+        nms_iou_threshold=self._nms_iou_threshold,
+        nms_max_output_boxes=self._nms_max_output_boxes)
     self._image_ids = set([])
     self._evaluate_corlocs = evaluate_corlocs
     self._evaluate_precision_recall = evaluate_precision_recall
@@ -454,7 +462,10 @@ class ObjectDetectionEvaluator(DetectionEvaluator):
         num_groundtruth_classes=self._num_classes,
         matching_iou_threshold=self._matching_iou_threshold,
         use_weighted_mean_ap=self._use_weighted_mean_ap,
-        label_id_offset=self._label_id_offset)
+        label_id_offset=self._label_id_offset,
+        nms_iou_threshold=self._nms_iou_threshold,
+        nms_max_output_boxes=self._nms_max_output_boxes,
+    )
     self._image_ids.clear()
 
   def add_eval_dict(self, eval_dict):
@@ -549,13 +560,19 @@ class ObjectDetectionEvaluator(DetectionEvaluator):
 class PascalDetectionEvaluator(ObjectDetectionEvaluator):
   """A class to evaluate detections using PASCAL metrics."""
 
-  def __init__(self, categories, matching_iou_threshold=0.5):
+  def __init__(self,
+               categories,
+               matching_iou_threshold=0.5,
+               nms_iou_threshold=1.0,
+               nms_max_output_boxes=10000):
     super(PascalDetectionEvaluator, self).__init__(
         categories,
         matching_iou_threshold=matching_iou_threshold,
         evaluate_corlocs=False,
         metric_prefix='PascalBoxes',
-        use_weighted_mean_ap=False)
+        use_weighted_mean_ap=False,
+        nms_iou_threshold=nms_iou_threshold,
+        nms_max_output_boxes=nms_max_output_boxes)
 
 
 class WeightedPascalDetectionEvaluator(ObjectDetectionEvaluator):