diff --git a/visualDet3D/networks/heads/detection_3d_head.py b/visualDet3D/networks/heads/detection_3d_head.py
index 8e66da5..6e950b2 100644
--- a/visualDet3D/networks/heads/detection_3d_head.py
+++ b/visualDet3D/networks/heads/detection_3d_head.py
@@ -465,10 +465,13 @@ def loss(self, cls_scores, reg_preds, anchors, annotations, P2s):
                     pos_anchor = anchor[pos_inds]
                     pos_alpha_score = alpha_score[pos_inds]
                     if self.decode_before_loss:
-                        pos_prediction_decoded = self._decode(pos_anchor, reg_pred[pos_inds],  anchors_3d_mean_std, label_index, pos_alpha_score)
-                        pos_target_decoded     = self._decode(pos_anchor, pos_bbox_targets,  anchors_3d_mean_std, label_index, pos_alpha_score)
-
-                        reg_loss.append((self.loss_bbox(pos_prediction_decoded, pos_target_decoded)* self.regression_weight).mean(dim=0))
+                        pos_prediction_decoded, mask = self._decode(pos_anchor, reg_pred[pos_inds],  anchor_mean_std_3d_j[pos_inds], label_index, pos_alpha_score)
+                        pos_target_decoded, _     = self._decode(pos_anchor, pos_bbox_targets,  anchor_mean_std_3d_j[pos_inds], label_index, pos_alpha_score)
+                        reg_loss_j = self.loss_bbox(pos_prediction_decoded[mask], pos_target_decoded[mask])
+                        alpha_loss_j = self.alpha_loss(pos_alpha_score[mask], targets_alpha_cls[mask])
+                        loss_j = torch.cat([reg_loss_j, alpha_loss_j], dim=1) * self.regression_weight #[N, 12]
+                        reg_loss.append(loss_j.mean(dim=0)) #[13]
+                        number_of_positives.append(bbox_annotation.shape[0])
                     else:
                         reg_loss_j = self.loss_bbox(pos_bbox_targets, reg_pred[pos_inds]) 
                         alpha_loss_j = self.alpha_loss(pos_alpha_score, targets_alpha_cls)