diff --git a/rcnn/BatchedNmsPlugin.h b/rcnn/BatchedNmsPlugin.h
index 476a67b..49467e8 100644
--- a/rcnn/BatchedNmsPlugin.h
+++ b/rcnn/BatchedNmsPlugin.h
@@ -30,7 +30,7 @@ class BatchedNmsPlugin : public IPluginV2Ext {
     float _nms_thresh;
     int _detections_per_im;
 
-    size_t _count;
+    size_t _count = 1;
 
  protected:
     void deserialize(void const* data, size_t length) {
diff --git a/rcnn/README.md b/rcnn/README.md
index c4b4bb1..7ea3dd1 100644
--- a/rcnn/README.md
+++ b/rcnn/README.md
@@ -124,6 +124,8 @@ sudo ./rcnn -d faster.engine ../samples
 
 - you can build fasterRcnn with maskRcnn weights file.
 
+- do initializing for _pre_nms_topk in RpnNmsPlugin and _count in BatchedNmsPlugin inside class to prevent error assert, because the configurePlugin function is implemented after clone() and before serialize(). one can also set it through constructor.
+
 ## Quantization
 
 1. quantizationType:fp32,fp16,int8. see BuildRcnnModel(rcnn.cpp line 276) for detail.
diff --git a/rcnn/RpnNmsPlugin.h b/rcnn/RpnNmsPlugin.h
index 792e5e9..7deb18e 100644
--- a/rcnn/RpnNmsPlugin.h
+++ b/rcnn/RpnNmsPlugin.h
@@ -28,7 +28,7 @@ class RpnNmsPlugin : public IPluginV2Ext {
     float _nms_thresh;
     int _post_nms_topk;
 
-    size_t _pre_nms_topk;
+    size_t _pre_nms_topk = 1;
     mutable int size = -1;
 
  protected: