diff --git a/paddle/fluid/operators/detection/matrix_nms_op.cc b/paddle/fluid/operators/detection/matrix_nms_op.cc
index f7d45bc85bf6b14f07e9bfda3615e0b2d51a09f1..713c2dc7fe9c15916fba15dcfffe6c5f7e2a3958 100644
--- a/paddle/fluid/operators/detection/matrix_nms_op.cc
+++ b/paddle/fluid/operators/detection/matrix_nms_op.cc
@@ -12,6 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 limitations under the License. */
 
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/op_version_registry.h"
 #include "paddle/fluid/operators/detection/nms_util.h"
 
 namespace paddle {
@@ -59,6 +60,9 @@ class MatrixNMSOp : public framework::OperatorWithKernel {
     }
     ctx->SetOutputDim("Out", {box_dims[1], box_dims[2] + 2});
     ctx->SetOutputDim("Index", {box_dims[1], 1});
+    if (ctx->HasOutput("RoisNum")) {
+      ctx->SetOutputDim("RoisNum", {-1});
+    }
     if (!ctx->IsRuntime()) {
       ctx->SetLoDLevel("Out", std::max(ctx->GetLoDLevel("BBoxes"), 1));
       ctx->SetLoDLevel("Index", std::max(ctx->GetLoDLevel("BBoxes"), 1));
@@ -259,8 +263,10 @@ class MatrixNMSKernel : public framework::OpKernel<T> {
     std::vector<size_t> offsets = {0};
     std::vector<T> detections;
     std::vector<int> indices;
+    std::vector<int> num_per_batch;
     detections.reserve(out_dim * num_boxes * batch_size);
     indices.reserve(num_boxes * batch_size);
+    num_per_batch.reserve(batch_size);
     for (int i = 0; i < batch_size; ++i) {
       scores_slice = scores->Slice(i, i + 1);
       scores_slice.Resize({score_dims[1], score_dims[2]});
@@ -272,6 +278,7 @@ class MatrixNMSKernel : public framework::OpKernel<T> {
           background_label, nms_top_k, keep_top_k, normalized, score_threshold,
           post_threshold, use_gaussian, gaussian_sigma);
       offsets.push_back(offsets.back() + num_out);
+      num_per_batch.emplace_back(num_out);
     }
 
     int64_t num_kept = offsets.back();
@@ -285,6 +292,12 @@ class MatrixNMSKernel : public framework::OpKernel<T> {
       std::copy(indices.begin(), indices.end(), index->data<int>());
     }
 
+    if (ctx.HasOutput("RoisNum")) {
+      auto* rois_num = ctx.Output<Tensor>("RoisNum");
+      rois_num->mutable_data<int>({batch_size}, ctx.GetPlace());
+      std::copy(num_per_batch.begin(), num_per_batch.end(),
+                rois_num->data<int>());
+    }
     framework::LoD lod;
     lod.emplace_back(offsets);
     outs->set_lod(lod);
@@ -355,6 +368,8 @@ class MatrixNMSOpMaker : public framework::OpProtoAndCheckerMaker {
               "(LoDTensor) A 2-D LoDTensor with shape [No, 1] represents the "
               "index of selected bbox. The index is the absolute index cross "
               "batches.");
+    AddOutput("RoisNum", "(Tensor), Number of RoIs in each images.")
+        .AsDispensable();
     AddComment(R"DOC(
 This operator does multi-class matrix non maximum suppression (NMS) on batched
 boxes and scores.
@@ -369,7 +384,9 @@ This operator support multi-class and batched inputs. It applying NMS
 independently for each class. The outputs is a 2-D LoDTenosr, for each
 image, the offsets in first dimension of LoDTensor are called LoD, the number
 of offset is N + 1, where N is the batch size. If LoD[i + 1] - LoD[i] == 0,
-means there is no detected bbox for this image.
+means there is no detected bbox for this image. Now this operator has one more
+ouput, which is RoisNum. The size of RoisNum is N, RoisNum[i] means the number of 
+detected bbox for this image.
 
 For more information on Matrix NMS, please refer to:
 https://arxiv.org/abs/2003.10152
@@ -387,3 +404,8 @@ REGISTER_OPERATOR(
     paddle::framework::EmptyGradOpMaker<paddle::imperative::OpBase>);
 REGISTER_OP_CPU_KERNEL(matrix_nms, ops::MatrixNMSKernel<float>,
                        ops::MatrixNMSKernel<double>);
+REGISTER_OP_VERSION(matrix_nms)
+    .AddCheckpoint(
+        R"ROC(Upgrade matrix_nms: add a new output [RoisNum].)ROC",
+        paddle::framework::compatible::OpVersionDesc().NewOutput(
+            "RoisNum", "The number of RoIs in each image."));
diff --git a/paddle/fluid/pybind/op_function_generator.cc b/paddle/fluid/pybind/op_function_generator.cc
index 8288f1852c27b1cf256e0c171bb4bf962e6caf6b..7f2736a9b1d4149f54c8eb516d1884141c90d446 100644
--- a/paddle/fluid/pybind/op_function_generator.cc
+++ b/paddle/fluid/pybind/op_function_generator.cc
@@ -74,6 +74,7 @@ std::map<std::string, std::set<std::string>> op_outs_map = {
     {"unique", {"Out", "Index", "Indices", "Counts"}},
     {"generate_proposals", {"RpnRois", "RpnRoiProbs", "RpnRoisNum"}},
     {"collect_fpn_proposals", {"FpnRois", "RoisNum"}},
+    {"matrix_nms", {"Out", "Index", "RoisNum"}},
     {"distribute_fpn_proposals",
      {"MultiFpnRois", "RestoreIndex", "MultiLevelRoIsNum"}},
     {"moving_average_abs_max_scale", {"OutScale", "OutAccum", "OutState"}},
diff --git a/python/paddle/fluid/tests/unittests/test_matrix_nms_op.py b/python/paddle/fluid/tests/unittests/test_matrix_nms_op.py
index cf756ae8384486ea30ebc2fd25079484061ed380..2bbacc316f6e6095453a7875c78f6a0a4c677ebe 100644
--- a/python/paddle/fluid/tests/unittests/test_matrix_nms_op.py
+++ b/python/paddle/fluid/tests/unittests/test_matrix_nms_op.py
@@ -201,7 +201,8 @@ class TestMatrixNMSOp(OpTest):
         self.inputs = {'BBoxes': boxes, 'Scores': scores}
         self.outputs = {
             'Out': (nmsed_outs, [lod]),
-            'Index': (index_outs[:, None], [lod])
+            'Index': (index_outs[:, None], [lod]),
+            'RoisNum': np.array(lod).astype('int32')
         }
         self.attrs = {
             'background_label': 0,