diff --git a/configs/centernet/_base_/centernet_reader.yml b/configs/centernet/_base_/centernet_reader.yml
index 1f18dca49d1e39c61b9bbc7b5be8bfac7bce5ca4..81af4ab840502da6e738ac667dd0883041ba8992 100644
--- a/configs/centernet/_base_/centernet_reader.yml
+++ b/configs/centernet/_base_/centernet_reader.yml
@@ -30,6 +30,6 @@ TestReader:
   sample_transforms:
     - Decode: {}
     - WarpAffine: {keep_res: True, input_h: 512, input_w: 512}
-    - NormalizeImage: {mean: [0.40789655, 0.44719303, 0.47026116], std: [0.2886383 , 0.27408165, 0.27809834]}
+    - NormalizeImage: {mean: [0.40789655, 0.44719303, 0.47026116], std: [0.2886383 , 0.27408165, 0.27809834], is_scale: True}
     - Permute: {}
   batch_size: 1
diff --git a/deploy/cpp/include/preprocess_op.h b/deploy/cpp/include/preprocess_op.h
index 33d7300b8fd84287cca91e86214084acec781030..2d24799a332293a8ef14a32d6d09f6daaa116425 100644
--- a/deploy/cpp/include/preprocess_op.h
+++ b/deploy/cpp/include/preprocess_op.h
@@ -74,7 +74,7 @@ class NormalizeImage : public PreprocessOp {
   // CHW or HWC
   std::vector<float> mean_;
   std::vector<float> scale_;
-  bool is_scale_;
+  bool is_scale_ = true;
 };
 
 class Permute : public PreprocessOp {
@@ -143,6 +143,24 @@ class TopDownEvalAffine : public PreprocessOp {
   std::vector<int> trainsize_;
 };
 
+class WarpAffine : public PreprocessOp {
+ public:
+  virtual void Init(const YAML::Node& item) {
+    input_h_ = item["input_h"].as<int>();
+    input_w_ = item["input_w"].as<int>();
+    keep_res_ = item["keep_res"].as<bool>();
+  }
+
+  virtual void Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  int input_h_;
+  int input_w_;
+  int interp_ = 1;
+  bool keep_res_ = true;
+  int pad_ = 31;
+};
+
 void CropImg(cv::Mat& img,
              cv::Mat& crop_img,
              std::vector<int>& area,
@@ -183,6 +201,8 @@ class Preprocessor {
       return std::make_shared<PadStride>();
     } else if (name == "TopDownEvalAffine") {
       return std::make_shared<TopDownEvalAffine>();
+    } else if (name == "WarpAffine") {
+      return std::make_shared<WarpAffine>();
     }
     std::cerr << "can not find function of OP: " << name
               << " and return: nullptr" << std::endl;
diff --git a/deploy/cpp/src/preprocess_op.cc b/deploy/cpp/src/preprocess_op.cc
index d4a1fb4191b449267581c71e94b5ff959ac886c3..98c700f6219ca4201d87902d2b337d5adc4d0d5c 100644
--- a/deploy/cpp/src/preprocess_op.cc
+++ b/deploy/cpp/src/preprocess_op.cc
@@ -177,11 +177,64 @@ void TopDownEvalAffine::Run(cv::Mat* im, ImageBlob* data) {
   };
 }
 
+void GetAffineTrans(const cv::Point2f center,
+                    const cv::Point2f input_size,
+                    const cv::Point2f output_size,
+                    cv::Mat* trans) {
+  cv::Point2f srcTri[3];
+  cv::Point2f dstTri[3];
+  float src_w = input_size.x;
+  float dst_w = output_size.x;
+  float dst_h = output_size.y;
+
+  cv::Point2f src_dir(0, -0.5 * src_w);
+  cv::Point2f dst_dir(0, -0.5 * dst_w);
+
+  srcTri[0] = center;
+  srcTri[1] = center + src_dir;
+  cv::Point2f src_d = srcTri[0] - srcTri[1];
+  srcTri[2] = srcTri[1] + cv::Point2f(-src_d.y, src_d.x);
+
+  dstTri[0] = cv::Point2f(dst_w * 0.5, dst_h * 0.5);
+  dstTri[1] = cv::Point2f(dst_w * 0.5, dst_h * 0.5) + dst_dir;
+  cv::Point2f dst_d = dstTri[0] - dstTri[1];
+  dstTri[2] = dstTri[1] + cv::Point2f(-dst_d.y, dst_d.x);
+
+  *trans = cv::getAffineTransform(srcTri, dstTri);
+}
+
+void WarpAffine::Run(cv::Mat* im, ImageBlob* data) {
+  cv::cvtColor(*im, *im, cv::COLOR_RGB2BGR);
+  cv::Mat trans(2, 3, CV_32FC1);
+  cv::Point2f center;
+  cv::Point2f input_size;
+  int h = im->rows;
+  int w = im->cols;
+  if (keep_res_) {
+    input_h_ = (h | pad_) + 1;
+    input_w_ = (w + pad_) + 1;
+    input_size = cv::Point2f(input_w_, input_h_);
+    center = cv::Point2f(w / 2, h / 2);
+  } else {
+    float s = std::max(h, w) * 1.0;
+    input_size = cv::Point2f(s, s);
+    center = cv::Point2f(w / 2., h / 2.);
+  }
+  cv::Point2f output_size(input_w_, input_h_);
+
+  GetAffineTrans(center, input_size, output_size, &trans);
+  cv::warpAffine(*im, *im, trans, cv::Size(input_w_, input_h_));
+  data->in_net_shape_ = {
+      static_cast<float>(input_h_), static_cast<float>(input_w_),
+  };
+}
+
 // Preprocessor op running order
 const std::vector<std::string> Preprocessor::RUN_ORDER = {"InitInfo",
                                                           "TopDownEvalAffine",
                                                           "Resize",
                                                           "LetterBoxResize",
+                                                          "WarpAffine",
                                                           "NormalizeImage",
                                                           "PadStride",
                                                           "Permute"};
diff --git a/ppdet/modeling/post_process.py b/ppdet/modeling/post_process.py
index e74095505222142d47c17a0e665d358561558173..c64befa8176cd7fb9fb820a0aa3d82adbf7e4396 100644
--- a/ppdet/modeling/post_process.py
+++ b/ppdet/modeling/post_process.py
@@ -524,7 +524,7 @@ class CenterNetPostProcess(TTFBox):
             x2 = xs + wh[:, 0:1] / 2
             y2 = ys + wh[:, 1:2] / 2
 
-        n, c, feat_h, feat_w = hm.shape[:]
+        n, c, feat_h, feat_w = paddle.shape(hm)
         padw = (feat_w * self.down_ratio - im_shape[0, 1]) / 2
         padh = (feat_h * self.down_ratio - im_shape[0, 0]) / 2
         x1 = x1 * self.down_ratio