Slice groundtruth tensors to remove padding during evaluation.

PiperOrigin-RevId: 343491021

research/object_detection/metrics/coco_evaluation.py

@@ -373,8 +373,9 @@ class CocoDetectionEvaluator(object_detection_evaluation.DetectionEvaluator):
     detection_scores = eval_dict[detection_fields.detection_scores]
     detection_classes = eval_dict[detection_fields.detection_classes]
     num_gt_boxes_per_image = eval_dict.get(
-        'num_groundtruth_boxes_per_image', None)
-    num_det_boxes_per_image = eval_dict.get('num_det_boxes_per_image', None)
+        input_data_fields.num_groundtruth_boxes, None)
+    num_det_boxes_per_image = eval_dict.get(detection_fields.num_detections,
+                                            None)
     is_annotated = eval_dict.get('is_annotated', None)
 
     if groundtruth_is_crowd is None:

research/object_detection/model_lib_v2.py

@@ -890,8 +890,7 @@ def eager_eval_loop(
       tf.logging.info('Finished eval step %d', i)
 
     use_original_images = fields.InputDataFields.original_image in features
-    if (use_original_images and i < eval_config.num_visualizations
-        and batch_size == 1):
+    if (use_original_images and i < eval_config.num_visualizations):
       sbys_image_list = vutils.draw_side_by_side_evaluation_image(
           eval_dict,
           category_index=category_index,
@@ -899,20 +898,20 @@ def eager_eval_loop(
           min_score_thresh=eval_config.min_score_threshold,
           use_normalized_coordinates=False,
           keypoint_edges=keypoint_edges or None)
-      sbys_images = tf.concat(sbys_image_list, axis=0)
+      for j, sbys_image in enumerate(sbys_image_list):
         tf.compat.v2.summary.image(
-          name='eval_side_by_side_' + str(i),
+            name='eval_side_by_side_{}_{}'.format(i, j),
             step=global_step,
-          data=sbys_images,
+            data=sbys_image,
             max_outputs=eval_config.num_visualizations)
       if eval_util.has_densepose(eval_dict):
         dp_image_list = vutils.draw_densepose_visualizations(
             eval_dict)
-        dp_images = tf.concat(dp_image_list, axis=0)
+        for j, dp_image in enumerate(dp_image_list):
           tf.compat.v2.summary.image(
-            name='densepose_detections_' + str(i),
+              name='densepose_detections_{}_{}'.format(i, j),
               step=global_step,
-            data=dp_images,
+              data=dp_image,
               max_outputs=eval_config.num_visualizations)
 
     if evaluators is None:

research/object_detection/utils/visualization_utils.py

@@ -664,6 +664,10 @@ def draw_side_by_side_evaluation_image(eval_dict,
           key != input_data_fields.image_additional_channels):
         eval_dict[key] = tf.expand_dims(eval_dict[key], 0)
 
+  num_gt_boxes = [-1] * eval_dict[input_data_fields.original_image].shape[0]
+  if input_data_fields.num_groundtruth_boxes in eval_dict:
+    num_gt_boxes = tf.cast(eval_dict[input_data_fields.num_groundtruth_boxes],
+                           tf.int32)
   for indx in range(eval_dict[input_data_fields.original_image].shape[0]):
     instance_masks = None
     if detection_fields.detection_masks in eval_dict:
@@ -702,7 +706,6 @@ def draw_side_by_side_evaluation_image(eval_dict,
         groundtruth_keypoint_scores = tf.cast(
             keypoint_ops.set_keypoint_visibilities(
                 groundtruth_keypoints), dtype=tf.float32)
-
     images_with_detections = draw_bounding_boxes_on_image_tensors(
         tf.expand_dims(
             eval_dict[input_data_fields.original_image][indx], axis=0),
@@ -725,16 +728,23 @@ def draw_side_by_side_evaluation_image(eval_dict,
         max_boxes_to_draw=max_boxes_to_draw,
         min_score_thresh=min_score_thresh,
         use_normalized_coordinates=use_normalized_coordinates)
+    num_gt_boxes_i = num_gt_boxes[indx]
     images_with_groundtruth = draw_bounding_boxes_on_image_tensors(
         tf.expand_dims(
-            eval_dict[input_data_fields.original_image][indx], axis=0),
+            eval_dict[input_data_fields.original_image][indx],
+            axis=0),
         tf.expand_dims(
-            eval_dict[input_data_fields.groundtruth_boxes][indx], axis=0),
+            eval_dict[input_data_fields.groundtruth_boxes][indx]
+            [:num_gt_boxes_i],
+            axis=0),
         tf.expand_dims(
-            eval_dict[input_data_fields.groundtruth_classes][indx], axis=0),
+            eval_dict[input_data_fields.groundtruth_classes][indx]
+            [:num_gt_boxes_i],
+            axis=0),
         tf.expand_dims(
             tf.ones_like(
-                eval_dict[input_data_fields.groundtruth_classes][indx],
+                eval_dict[input_data_fields.groundtruth_classes][indx]
+                [:num_gt_boxes_i],
                 dtype=tf.float32),
             axis=0),
         category_index,
@@ -760,13 +770,17 @@ def draw_side_by_side_evaluation_image(eval_dict,
                   eval_dict[input_data_fields.image_additional_channels][indx],
                   axis=0),
               tf.expand_dims(
-                  eval_dict[input_data_fields.groundtruth_boxes][indx], axis=0),
+                  eval_dict[input_data_fields.groundtruth_boxes][indx]
+                  [:num_gt_boxes_i],
+                  axis=0),
               tf.expand_dims(
-                  eval_dict[input_data_fields.groundtruth_classes][indx],
+                  eval_dict[input_data_fields.groundtruth_classes][indx]
+                  [:num_gt_boxes_i],
                   axis=0),
               tf.expand_dims(
                   tf.ones_like(
-                      eval_dict[input_data_fields.groundtruth_classes][indx],
+                      eval_dict[input_data_fields.groundtruth_classes][indx]
+                      [num_gt_boxes_i],
                       dtype=tf.float32),
                   axis=0),
               category_index,