diff --git a/paddle/gserver/evaluators/DetectionMAPEvaluator.cpp b/paddle/gserver/evaluators/DetectionMAPEvaluator.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..7d326c2db144434605fa8d1c03c722dcdf010d23
--- /dev/null
+++ b/paddle/gserver/evaluators/DetectionMAPEvaluator.cpp
@@ -0,0 +1,312 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "Evaluator.h"
+#include "paddle/gserver/layers/DetectionUtil.h"
+
+using std::map;
+using std::vector;
+using std::pair;
+using std::make_pair;
+
+namespace paddle {
+
+/**
+ * @brief detection map Evaluator
+ *
+ * The config file api is detection_map_evaluator.
+ */
+class DetectionMAPEvaluator : public Evaluator {
+public:
+  DetectionMAPEvaluator()
+      : evaluateDifficult_(false), cpuOutput_(nullptr), cpuLabel_(nullptr) {}
+
+  virtual void start() {
+    Evaluator::start();
+    allTruePos_.clear();
+    allFalsePos_.clear();
+    numPos_.clear();
+  }
+
+  virtual real evalImp(std::vector<Argument>& arguments) {
+    overlapThreshold_ = config_.overlap_threshold();
+    backgroundId_ = config_.background_id();
+    evaluateDifficult_ = config_.evaluate_difficult();
+    apType_ = config_.ap_type();
+
+    MatrixPtr detectTmpValue = arguments[0].value;
+    Matrix::resizeOrCreate(cpuOutput_,
+                           detectTmpValue->getHeight(),
+                           detectTmpValue->getWidth(),
+                           false,
+                           false);
+
+    MatrixPtr labelTmpValue = arguments[1].value;
+    Matrix::resizeOrCreate(cpuLabel_,
+                           labelTmpValue->getHeight(),
+                           labelTmpValue->getWidth(),
+                           false,
+                           false);
+
+    cpuOutput_->copyFrom(*detectTmpValue);
+    cpuLabel_->copyFrom(*labelTmpValue);
+
+    Argument label = arguments[1];
+    const int* labelIndex = label.sequenceStartPositions->getData(false);
+    size_t batchSize = label.getNumSequences();
+
+    vector<map<size_t, vector<NormalizedBBox>>> allGTBBoxes;
+    vector<map<size_t, vector<pair<real, NormalizedBBox>>>> allDetectBBoxes;
+
+    for (size_t n = 0; n < batchSize; ++n) {
+      map<size_t, vector<NormalizedBBox>> bboxes;
+      for (int i = labelIndex[n]; i < labelIndex[n + 1]; ++i) {
+        vector<NormalizedBBox> bbox;
+        getBBoxFromLabelData(cpuLabel_->getData() + i * 6, 1, bbox);
+        int c = cpuLabel_->getData()[i * 6];
+        bboxes[c].push_back(bbox[0]);
+      }
+      allGTBBoxes.push_back(bboxes);
+    }
+
+    size_t imgId = 0;
+    for (size_t n = 0; n < cpuOutput_->getHeight();) {
+      map<size_t, vector<pair<real, NormalizedBBox>>> bboxes;
+      while (cpuOutput_->getData()[n * 7] == imgId &&
+             n < cpuOutput_->getHeight()) {
+        vector<real> label;
+        vector<real> score;
+        vector<NormalizedBBox> bbox;
+        getBBoxFromDetectData(
+            cpuOutput_->getData() + n * 7, 1, label, score, bbox);
+        bboxes[label[0]].push_back(make_pair(score[0], bbox[0]));
+        ++n;
+      }
+      ++imgId;
+      if (imgId > batchSize) break;
+      allDetectBBoxes.push_back(bboxes);
+    }
+
+    for (size_t n = 0; n < batchSize; ++n) {
+      for (map<size_t, vector<NormalizedBBox>>::iterator it =
+               allGTBBoxes[n].begin();
+           it != allGTBBoxes[n].end();
+           ++it) {
+        size_t count = 0;
+        if (evaluateDifficult_) {
+          count = it->second.size();
+        } else {
+          for (size_t i = 0; i < it->second.size(); ++i)
+            if (!(it->second[i].isDifficult)) ++count;
+        }
+        if (numPos_.find(it->first) == numPos_.end() && count != 0) {
+          numPos_[it->first] = count;
+        } else {
+          numPos_[it->first] += count;
+        }
+      }
+    }
+
+    // calcTFPos
+    calcTFPos(
+        batchSize, allGTBBoxes, allDetectBBoxes, &allTruePos_, &allFalsePos_);
+
+    return 0;
+  }
+
+  virtual void printStats(std::ostream& os) const {
+    real mAP = calcMAP();
+    os << "Detection mAP=" << mAP * 100;
+  }
+
+  virtual void distributeEval(ParameterClient2* client) {
+    LOG(FATAL) << "Distribute detection evaluation not implemented.";
+  }
+
+protected:
+  void calcTFPos(const size_t batchSize,
+                 const vector<map<size_t, vector<NormalizedBBox>>>& allGTBBoxes,
+                 const vector<map<size_t, vector<pair<real, NormalizedBBox>>>>&
+                     allDetectBBoxes,
+                 map<size_t, vector<pair<real, size_t>>>* allTruePos,
+                 map<size_t, vector<pair<real, size_t>>>* allFalsePos) {
+    for (size_t n = 0; n < allDetectBBoxes.size(); ++n) {
+      if (allGTBBoxes[n].size() == 0) {
+        for (map<size_t, vector<pair<real, NormalizedBBox>>>::const_iterator
+                 it = allDetectBBoxes[n].begin();
+             it != allDetectBBoxes[n].end();
+             ++it) {
+          size_t label = it->first;
+          for (size_t i = 0; i < it->second.size(); ++i) {
+            (*allTruePos)[label].push_back(make_pair(it->second[i].first, 0));
+            (*allFalsePos)[label].push_back(make_pair(it->second[i].first, 1));
+          }
+        }
+      } else {
+        for (map<size_t, vector<pair<real, NormalizedBBox>>>::const_iterator
+                 it = allDetectBBoxes[n].begin();
+             it != allDetectBBoxes[n].end();
+             ++it) {
+          size_t label = it->first;
+          vector<pair<real, NormalizedBBox>> predBBoxes = it->second;
+          if (allGTBBoxes[n].find(label) == allGTBBoxes[n].end()) {
+            for (size_t i = 0; i < predBBoxes.size(); ++i) {
+              (*allTruePos)[label].push_back(make_pair(predBBoxes[i].first, 0));
+              (*allFalsePos)[label].push_back(
+                  make_pair(predBBoxes[i].first, 1));
+            }
+          } else {
+            vector<NormalizedBBox> gtBBoxes =
+                allGTBBoxes[n].find(label)->second;
+            vector<bool> visited(gtBBoxes.size(), false);
+            // Sort detections in descend order based on scores
+            std::sort(predBBoxes.begin(),
+                      predBBoxes.end(),
+                      sortScorePairDescend<NormalizedBBox>);
+            for (size_t i = 0; i < predBBoxes.size(); ++i) {
+              real maxOverlap = -1.0;
+              size_t maxIdx = 0;
+              for (size_t j = 0; j < gtBBoxes.size(); ++j) {
+                real overlap =
+                    jaccardOverlap(predBBoxes[i].second, gtBBoxes[j]);
+                if (overlap > maxOverlap) {
+                  maxOverlap = overlap;
+                  maxIdx = j;
+                }
+              }
+              if (maxOverlap > overlapThreshold_) {
+                if (evaluateDifficult_ ||
+                    (!evaluateDifficult_ && !gtBBoxes[maxIdx].isDifficult)) {
+                  if (!visited[maxIdx]) {
+                    (*allTruePos)[label].push_back(
+                        make_pair(predBBoxes[i].first, 1));
+                    (*allFalsePos)[label].push_back(
+                        make_pair(predBBoxes[i].first, 0));
+                    visited[maxIdx] = true;
+                  } else {
+                    (*allTruePos)[label].push_back(
+                        make_pair(predBBoxes[i].first, 0));
+                    (*allFalsePos)[label].push_back(
+                        make_pair(predBBoxes[i].first, 1));
+                  }
+                }
+              } else {
+                (*allTruePos)[label].push_back(
+                    make_pair(predBBoxes[i].first, 0));
+                (*allFalsePos)[label].push_back(
+                    make_pair(predBBoxes[i].first, 1));
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  real calcMAP() const {
+    real mAP = 0.0;
+    size_t count = 0;
+    for (map<size_t, size_t>::const_iterator it = numPos_.begin();
+         it != numPos_.end();
+         ++it) {
+      size_t label = it->first;
+      size_t labelNumPos = it->second;
+      if (labelNumPos == 0 || allTruePos_.find(label) == allTruePos_.end())
+        continue;
+      vector<pair<real, size_t>> labelTruePos = allTruePos_.find(label)->second;
+      vector<pair<real, size_t>> labelFalsePos =
+          allFalsePos_.find(label)->second;
+      // Compute average precision.
+      vector<size_t> tpCumSum;
+      getAccumulation(labelTruePos, &tpCumSum);
+      vector<size_t> fpCumSum;
+      getAccumulation(labelFalsePos, &fpCumSum);
+      std::vector<real> precision, recall;
+      size_t num = tpCumSum.size();
+      // Compute Precision.
+      for (size_t i = 0; i < num; ++i) {
+        CHECK_LE(tpCumSum[i], labelNumPos);
+        precision.push_back(static_cast<real>(tpCumSum[i]) /
+                            static_cast<real>(tpCumSum[i] + fpCumSum[i]));
+        recall.push_back(static_cast<real>(tpCumSum[i]) / labelNumPos);
+      }
+      // VOC2007 style
+      if (apType_ == "11point") {
+        vector<real> maxPrecisions(11, 0.0);
+        int startIdx = num - 1;
+        for (int j = 10; j >= 0; --j)
+          for (int i = startIdx; i >= 0; --i) {
+            if (recall[i] < j / 10.) {
+              startIdx = i;
+              if (j > 0) maxPrecisions[j - 1] = maxPrecisions[j];
+              break;
+            } else {
+              if (maxPrecisions[j] < precision[i])
+                maxPrecisions[j] = precision[i];
+            }
+          }
+        for (int j = 10; j >= 0; --j) mAP += maxPrecisions[j] / 11;
+        ++count;
+      } else if (apType_ == "Integral") {
+        // Nature integral
+        real averagePrecisions = 0.;
+        real prevRecall = 0.;
+        for (size_t i = 0; i < num; ++i) {
+          if (fabs(recall[i] - prevRecall) > 1e-6)
+            averagePrecisions += precision[i] * fabs(recall[i] - prevRecall);
+          prevRecall = recall[i];
+        }
+        mAP += averagePrecisions;
+        ++count;
+      } else {
+        LOG(FATAL) << "Unkown ap version: " << apType_;
+      }
+    }
+    if (count != 0) mAP /= count;
+    return mAP;
+  }
+
+  void getAccumulation(vector<pair<real, size_t>> inPairs,
+                       vector<size_t>* accuVec) const {
+    std::stable_sort(
+        inPairs.begin(), inPairs.end(), sortScorePairDescend<size_t>);
+    accuVec->clear();
+    size_t sum = 0;
+    for (size_t i = 0; i < inPairs.size(); ++i) {
+      sum += inPairs[i].second;
+      accuVec->push_back(sum);
+    }
+  }
+
+  std::string getTypeImpl() const { return "detection_map"; }
+
+  real getValueImpl() const { return calcMAP() * 100; }
+
+private:
+  real overlapThreshold_;
+  bool evaluateDifficult_;
+  size_t backgroundId_;
+  std::string apType_;
+
+  MatrixPtr cpuOutput_;
+  MatrixPtr cpuLabel_;
+
+  map<size_t, size_t> numPos_;
+  map<size_t, vector<pair<real, size_t>>> allTruePos_;
+  map<size_t, vector<pair<real, size_t>>> allFalsePos_;
+};
+
+REGISTER_EVALUATOR(detection_map, DetectionMAPEvaluator);
+
+}  // namespace paddle
diff --git a/paddle/gserver/tests/test_Evaluator.cpp b/paddle/gserver/tests/test_Evaluator.cpp
index 4f5fdbb37ce024e18b8d39c5dda74c69bf82166a..93996392d221d531f65caf465decbffdbc2d0384 100644
--- a/paddle/gserver/tests/test_Evaluator.cpp
+++ b/paddle/gserver/tests/test_Evaluator.cpp
@@ -138,6 +138,23 @@ void testEvaluatorAll(TestConfig testConf,
   testEvaluator(testConf, testEvaluatorName, batchSize, false);
 }
 
+TEST(Evaluator, detection_map) {
+  TestConfig config;
+  config.evaluatorConfig.set_type("detection_map");
+  config.evaluatorConfig.set_overlap_threshold(0.5);
+  config.evaluatorConfig.set_background_id(0);
+  config.evaluatorConfig.set_ap_type("Integral");
+  config.evaluatorConfig.set_evaluate_difficult(0);
+
+  config.inputDefs.push_back({INPUT_DATA, "output", 7});
+  config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "label", 6});
+  config.evaluatorConfig.set_evaluate_difficult(false);
+  testEvaluatorAll(config, "detection_map", 100);
+
+  config.evaluatorConfig.set_evaluate_difficult(true);
+  testEvaluatorAll(config, "detection_map", 100);
+}
+
 TEST(Evaluator, classification_error) {
   TestConfig config;
   config.evaluatorConfig.set_type("classification_error");
diff --git a/proto/ModelConfig.proto b/proto/ModelConfig.proto
index 29270829bbc3af6990aaf03a5228ef7f6a892a5c..ebe4f5cbb569ff37a46eb44de6362a7df337fe38 100644
--- a/proto/ModelConfig.proto
+++ b/proto/ModelConfig.proto
@@ -489,6 +489,15 @@ message EvaluatorConfig {
   // Used by ClassificationErrorEvaluator
   // top # classification error
   optional int32 top_k = 13 [default = 1];
+
+  // Used by DetectionMAPEvaluator
+  optional double overlap_threshold = 14 [default = 0.5];
+
+  optional int32 background_id = 15 [default = 0];
+
+  optional bool evaluate_difficult = 16 [default = false];
+
+  optional string ap_type = 17 [default = "11point"];
 }
 
 message LinkConfig {
diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py
index 0792e2d40b43f5fb2de8d6bb43a62cfa23f77082..e78dc4f3b4d8501b9864ed6c7e38afa98720deb2 100644
--- a/python/paddle/trainer/config_parser.py
+++ b/python/paddle/trainer/config_parser.py
@@ -1300,20 +1300,23 @@ def parse_maxout(maxout, input_layer_name, maxout_conf):
 
 # Define an evaluator
 @config_func
-def Evaluator(
-        name,
-        type,
-        inputs,
-        chunk_scheme=None,
-        num_chunk_types=None,
-        classification_threshold=None,
-        positive_label=None,
-        dict_file=None,
-        result_file=None,
-        num_results=None,
-        top_k=None,
-        delimited=None,
-        excluded_chunk_types=None, ):
+def Evaluator(name,
+              type,
+              inputs,
+              chunk_scheme=None,
+              num_chunk_types=None,
+              classification_threshold=None,
+              positive_label=None,
+              dict_file=None,
+              result_file=None,
+              num_results=None,
+              top_k=None,
+              delimited=None,
+              excluded_chunk_types=None,
+              overlap_threshold=None,
+              background_id=None,
+              evaluate_difficult=None,
+              ap_type=None):
     evaluator = g_config.model_config.evaluators.add()
     evaluator.type = type
     evaluator.name = MakeLayerNameInSubmodel(name)
@@ -1347,6 +1350,18 @@ def Evaluator(
     if excluded_chunk_types:
         evaluator.excluded_chunk_types.extend(excluded_chunk_types)
 
+    if overlap_threshold is not None:
+        evaluator.overlap_threshold = overlap_threshold
+
+    if background_id is not None:
+        evaluator.background_id = background_id
+
+    if evaluate_difficult is not None:
+        evaluator.evaluate_difficult = evaluate_difficult
+
+    if ap_type is not None:
+        evaluator.ap_type = ap_type
+
 
 class LayerBase(object):
     def __init__(
diff --git a/python/paddle/trainer_config_helpers/evaluators.py b/python/paddle/trainer_config_helpers/evaluators.py
index a5234f3e47f6caa4b365de593648e0ee5ad6e4a2..1dcd80480325558c16710d195dc2e05aa6200b24 100644
--- a/python/paddle/trainer_config_helpers/evaluators.py
+++ b/python/paddle/trainer_config_helpers/evaluators.py
@@ -21,7 +21,8 @@ __all__ = [
     "chunk_evaluator", "sum_evaluator", "column_sum_evaluator",
     "value_printer_evaluator", "gradient_printer_evaluator",
     "maxid_printer_evaluator", "maxframe_printer_evaluator",
-    "seqtext_printer_evaluator", "classification_error_printer_evaluator"
+    "seqtext_printer_evaluator", "classification_error_printer_evaluator",
+    "detection_map_evaluator"
 ]
 
 
@@ -31,10 +32,11 @@ class EvaluatorAttribute(object):
     FOR_RANK = 1 << 2
     FOR_PRINT = 1 << 3
     FOR_UTILS = 1 << 4
+    FOR_DETECTION = 1 << 5
 
     KEYS = [
         "for_classification", "for_regression", "for_rank", "for_print",
-        "for_utils"
+        "for_utils", "for_detection"
     ]
 
     @staticmethod
@@ -57,22 +59,25 @@ def evaluator(*attrs):
     return impl
 
 
-def evaluator_base(
-        input,
-        type,
-        label=None,
-        weight=None,
-        name=None,
-        chunk_scheme=None,
-        num_chunk_types=None,
-        classification_threshold=None,
-        positive_label=None,
-        dict_file=None,
-        result_file=None,
-        num_results=None,
-        delimited=None,
-        top_k=None,
-        excluded_chunk_types=None, ):
+def evaluator_base(input,
+                   type,
+                   label=None,
+                   weight=None,
+                   name=None,
+                   chunk_scheme=None,
+                   num_chunk_types=None,
+                   classification_threshold=None,
+                   positive_label=None,
+                   dict_file=None,
+                   result_file=None,
+                   num_results=None,
+                   delimited=None,
+                   top_k=None,
+                   excluded_chunk_types=None,
+                   overlap_threshold=None,
+                   background_id=None,
+                   evaluate_difficult=None,
+                   ap_type=None):
     """
     Evaluator will evaluate the network status while training/testing.
 
@@ -107,6 +112,14 @@ def evaluator_base(
     :type weight: LayerOutput.
     :param top_k: number k in top-k error rate
     :type top_k: int
+    :param overlap_threshold: In detection tasks to filter detection results
+    :type overlap_threshold: float
+    :param background_id: Identifier of background class
+    :type background_id: int
+    :param evaluate_difficult: Whether to evaluate difficult objects
+    :type evaluate_difficult: bool
+    :param ap_type: How to calculate average persicion
+    :type ap_type: str
     """
     # inputs type assertions.
     assert classification_threshold is None or isinstance(
@@ -136,7 +149,61 @@ def evaluator_base(
         delimited=delimited,
         num_results=num_results,
         top_k=top_k,
-        excluded_chunk_types=excluded_chunk_types, )
+        excluded_chunk_types=excluded_chunk_types,
+        overlap_threshold=overlap_threshold,
+        background_id=background_id,
+        evaluate_difficult=evaluate_difficult,
+        ap_type=ap_type)
+
+
+@evaluator(EvaluatorAttribute.FOR_DETECTION)
+@wrap_name_default()
+def detection_map_evaluator(input,
+                            label,
+                            overlap_threshold=0.5,
+                            background_id=0,
+                            evaluate_difficult=False,
+                            ap_type="11point",
+                            name=None):
+    """
+    Detection mAP Evaluator. It will print mean Average Precision for detection.
+
+    The detection mAP Evaluator according to the detection_output's output count
+    the true positive and the false positive bbox and integral them to get the
+    mAP.
+
+    The simple usage is:
+
+    .. code-block:: python
+
+       eval =  detection_map_evaluator(input=det_output,label=lbl)
+
+    :param input: Input layer.
+    :type input: LayerOutput
+    :param label: Label layer.
+    :type label: LayerOutput
+    :param overlap_threshold: The bbox overlap threshold of a true positive.
+    :type overlap_threshold: float
+    :param background_id: The background class index.
+    :type background_id: int
+    :param evaluate_difficult: Wether evaluate a difficult ground truth.
+    :type evaluate_difficult: bool
+    """
+    if not isinstance(input, list):
+        input = [input]
+
+    if label:
+        input.append(label)
+
+    evaluator_base(
+        name=name,
+        type="detection_map",
+        input=input,
+        label=label,
+        overlap_threshold=overlap_threshold,
+        background_id=background_id,
+        evaluate_difficult=evaluate_difficult,
+        ap_type=ap_type)
 
 
 @evaluator(EvaluatorAttribute.FOR_CLASSIFICATION)