[Cascade R-CNN] only train with nonempty boxes

examples/FasterRCNN/modeling/model_cascade.py

@@ -3,7 +3,7 @@ import tensorflow as tf
 from tensorpack.tfutils import get_current_tower_context
 
 from config import config as cfg
-from utils.box_ops import pairwise_iou
+from utils.box_ops import pairwise_iou, area as tf_area
 from .model_box import clip_boxes
 from .model_frcnn import BoxProposals, FastRCNNHead, fastrcnn_outputs
 
@@ -27,8 +27,8 @@ class CascadeRCNNHead(object):
 
         self.num_cascade_stages = len(cfg.CASCADE.IOUS)
 
-        self.is_training = get_current_tower_context().is_training
-        if self.is_training:
+        self.training = get_current_tower_context().is_training
+        if self.training:
             @tf.custom_gradient
             def scale_gradient(x):
                 return x, lambda dy: dy * (1.0 / self.num_cascade_stages)
@@ -72,6 +72,8 @@ class CascadeRCNNHead(object):
 
         refined_boxes = head.decoded_output_boxes_class_agnostic()
         refined_boxes = clip_boxes(refined_boxes, self.image_shape2d)
+        if self.training:
+            refined_boxes = tf.boolean_mask(refined_boxes, tf_area(refined_boxes) > 0)
         return head, tf.stop_gradient(refined_boxes, name='output_boxes')
 
     def match_box_with_gt(self, boxes, iou_threshold):
@@ -81,7 +83,7 @@ class CascadeRCNNHead(object):
         Returns:
             BoxProposals
         """
-        if self.is_training:
+        if self.training:
             with tf.name_scope('match_box_with_gt_{}'.format(iou_threshold)):
                 iou = pairwise_iou(boxes, self.gt_boxes)  # NxM
                 max_iou_per_box = tf.reduce_max(iou, axis=1)  # N