Merge pull request #15418 from jerrywgz/refine_nms

Refine nms

paddle/fluid/API.spec

@@ -325,6 +325,7 @@ paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None
 paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name'], varargs=None, keywords=None, defaults=('encode_center_size', True, None))
 paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'class_num', 'ignore_thresh', 'loss_weight_xy', 'loss_weight_wh', 'loss_weight_conf_target', 'loss_weight_conf_notarget', 'loss_weight_class', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None))
+paddle.fluid.layers.multiclass_nms ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None))
 paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None))
 paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1))
 paddle.fluid.layers.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))

paddle/fluid/operators/detection/multiclass_nms_op.cc

@@ -9,9 +9,9 @@ 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.
-
 limitations under the License. */
 
+#include <glog/logging.h>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detection/poly_util.h"
 
@@ -35,30 +35,45 @@ class MultiClassNMSOp : public framework::OperatorWithKernel {
 
     auto box_dims = ctx->GetInputDim("BBoxes");
     auto score_dims = ctx->GetInputDim("Scores");
+    auto score_size = score_dims.size();
 
     if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE(score_size == 2 || score_size == 3,
+                     "The rank of Input(Scores) must be 2 or 3");
       PADDLE_ENFORCE_EQ(box_dims.size(), 3,
-                        "The rank of Input(BBoxes) must be 3.");
-      PADDLE_ENFORCE_EQ(score_dims.size(), 3,
-                        "The rank of Input(Scores) must be 3.");
-      PADDLE_ENFORCE(box_dims[2] == 4 || box_dims[2] == 8 ||
-                         box_dims[2] == 16 || box_dims[2] == 24 ||
-                         box_dims[2] == 32,
-                     "The 2nd dimension of Input(BBoxes) must be 4 or 8, "
-                     "represents the layout of coordinate "
-                     "[xmin, ymin, xmax, ymax] or "
-                     "4 points: [x1, y1, x2, y2, x3, y3, x4, y4] or "
-                     "8 points: [xi, yi] i= 1,2,...,8 or "
-                     "12 points: [xi, yi] i= 1,2,...,12 or "
-                     "16 points: [xi, yi] i= 1,2,...,16");
-      PADDLE_ENFORCE_EQ(box_dims[1], score_dims[2],
-                        "The 1st dimensiong of Input(BBoxes) must be equal to "
-                        "3rd dimension of Input(Scores), which represents the "
-                        "predicted bboxes.");
+                        "The rank of Input(BBoxes) must be 3");
+      if (score_size == 3) {
+        PADDLE_ENFORCE(box_dims[2] == 4 || box_dims[2] == 8 ||
+                           box_dims[2] == 16 || box_dims[2] == 24 ||
+                           box_dims[2] == 32,
+                       "The last dimension of Input(BBoxes) must be 4 or 8, "
+                       "represents the layout of coordinate "
+                       "[xmin, ymin, xmax, ymax] or "
+                       "4 points: [x1, y1, x2, y2, x3, y3, x4, y4] or "
+                       "8 points: [xi, yi] i= 1,2,...,8 or "
+                       "12 points: [xi, yi] i= 1,2,...,12 or "
+                       "16 points: [xi, yi] i= 1,2,...,16");
+        PADDLE_ENFORCE_EQ(
+            box_dims[1], score_dims[2],
+            "The 2nd dimension of Input(BBoxes) must be equal to "
+            "last dimension of Input(Scores), which represents the "
+            "predicted bboxes.");
+      } else {
+        PADDLE_ENFORCE(box_dims[2] == 4,
+                       "The last dimension of Input(BBoxes) must be 4");
+        PADDLE_ENFORCE_EQ(box_dims[1], score_dims[1],
+                          "The 2nd dimension of Input(BBoxes)"
+                          "must be equal to the 2nd dimension"
+                          " of Input(Scores)");
+      }
     }
     // Here the box_dims[0] is not the real dimension of output.
     // It will be rewritten in the computing kernel.
-    ctx->SetOutputDim("Out", {box_dims[1], box_dims[2] + 2});
+    if (score_size == 3) {
+      ctx->SetOutputDim("Out", {box_dims[1], box_dims[2] + 2});
+    } else {
+      ctx->SetOutputDim("Out", {-1, box_dims[2] + 2});
+    }
   }
 
  protected:
@@ -123,8 +138,9 @@ static inline T JaccardOverlap(const T* box1, const T* box2,
     const T inter_ymin = std::max(box1[1], box2[1]);
     const T inter_xmax = std::min(box1[2], box2[2]);
     const T inter_ymax = std::min(box1[3], box2[3]);
-    const T inter_w = inter_xmax - inter_xmin;
-    const T inter_h = inter_ymax - inter_ymin;
+    T norm = normalized ? static_cast<T>(0.) : static_cast<T>(1.);
+    T inter_w = inter_xmax - inter_xmin + norm;
+    T inter_h = inter_ymax - inter_ymin + norm;
     const T inter_area = inter_w * inter_h;
     const T bbox1_area = BBoxArea<T>(box1, normalized);
     const T bbox2_area = BBoxArea<T>(box2, normalized);
@@ -139,7 +155,7 @@ T PolyIoU(const T* box1, const T* box2, const size_t box_size,
   T bbox2_area = PolyArea<T>(box2, box_size, normalized);
   T inter_area = PolyOverlapArea<T>(box1, box2, box_size, normalized);
   if (bbox1_area == 0 || bbox2_area == 0 || inter_area == 0) {
-    // If coordinate values are is invalid
+    // If coordinate values are invalid
     // if area size <= 0,  return 0.
     return T(0.);
   } else {
@@ -147,12 +163,35 @@ T PolyIoU(const T* box1, const T* box2, const size_t box_size,
   }
 }
 
+template <class T>
+void SliceOneClass(const platform::DeviceContext& ctx,
+                   const framework::Tensor& items, const int class_id,
+                   framework::Tensor* one_class_item) {
+  T* item_data = one_class_item->mutable_data<T>(ctx.GetPlace());
+  const T* items_data = items.data<T>();
+  const int64_t num_item = items.dims()[0];
+  const int class_num = items.dims()[1];
+  if (items.dims().size() == 3) {
+    int item_size = items.dims()[2];
+    for (int i = 0; i < num_item; ++i) {
+      std::memcpy(item_data + i * item_size,
+                  items_data + i * class_num * item_size + class_id * item_size,
+                  sizeof(T) * item_size);
+    }
+  } else {
+    for (int i = 0; i < num_item; ++i) {
+      item_data[i] = items_data[i * class_num + class_id];
+    }
+  }
+}
+
 template <typename T>
 class MultiClassNMSKernel : public framework::OpKernel<T> {
  public:
   void NMSFast(const Tensor& bbox, const Tensor& scores,
                const T score_threshold, const T nms_threshold, const T eta,
-               const int64_t top_k, std::vector<int>* selected_indices) const {
+               const int64_t top_k, std::vector<int>* selected_indices,
+               const bool normalized) const {
     // The total boxes for each instance.
     int64_t num_boxes = bbox.dims()[0];
     // 4: [xmin ymin xmax ymax]
@@ -178,15 +217,16 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
           T overlap = T(0.);
           // 4: [xmin ymin xmax ymax]
           if (box_size == 4) {
-            overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
-                                        bbox_data + kept_idx * box_size, true);
+            overlap =
+                JaccardOverlap<T>(bbox_data + idx * box_size,
+                                  bbox_data + kept_idx * box_size, normalized);
           }
           // 8: [x1 y1 x2 y2 x3 y3 x4 y4] or 16, 24, 32
           if (box_size == 8 || box_size == 16 || box_size == 24 ||
               box_size == 32) {
-            overlap =
-                PolyIoU<T>(bbox_data + idx * box_size,
-                           bbox_data + kept_idx * box_size, box_size, true);
+            overlap = PolyIoU<T>(bbox_data + idx * box_size,
+                                 bbox_data + kept_idx * box_size, box_size,
+                                 normalized);
           }
           keep = overlap <= adaptive_threshold;
         } else {
@@ -205,37 +245,58 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
 
   void MultiClassNMS(const framework::ExecutionContext& ctx,
                      const Tensor& scores, const Tensor& bboxes,
+                     const int scores_size,
                      std::map<int, std::vector<int>>* indices,
                      int* num_nmsed_out) const {
     int64_t background_label = ctx.Attr<int>("background_label");
     int64_t nms_top_k = ctx.Attr<int>("nms_top_k");
     int64_t keep_top_k = ctx.Attr<int>("keep_top_k");
+    bool normalized = ctx.Attr<bool>("normalized");
     T nms_threshold = static_cast<T>(ctx.Attr<float>("nms_threshold"));
     T nms_eta = static_cast<T>(ctx.Attr<float>("nms_eta"));
     T score_threshold = static_cast<T>(ctx.Attr<float>("score_threshold"));
+    auto& dev_ctx = ctx.template device_context<platform::CPUDeviceContext>();
 
-    int64_t class_num = scores.dims()[0];
-    int64_t predict_dim = scores.dims()[1];
     int num_det = 0;
+
+    int64_t class_num = scores_size == 3 ? scores.dims()[0] : scores.dims()[1];
+    Tensor bbox_slice, score_slice;
     for (int64_t c = 0; c < class_num; ++c) {
       if (c == background_label) continue;
-      Tensor score = scores.Slice(c, c + 1);
-      NMSFast(bboxes, score, score_threshold, nms_threshold, nms_eta, nms_top_k,
-              &((*indices)[c]));
+      if (scores_size == 3) {
+        score_slice = scores.Slice(c, c + 1);
+        bbox_slice = bboxes;
+      } else {
+        score_slice.Resize({scores.dims()[0], 1});
+        bbox_slice.Resize({scores.dims()[0], 4});
+        SliceOneClass<T>(dev_ctx, scores, c, &score_slice);
+        SliceOneClass<T>(dev_ctx, bboxes, c, &bbox_slice);
+      }
+      NMSFast(bbox_slice, score_slice, score_threshold, nms_threshold, nms_eta,
+              nms_top_k, &((*indices)[c]), normalized);
+      if (scores_size == 2) {
+        std::stable_sort((*indices)[c].begin(), (*indices)[c].end());
+      }
       num_det += (*indices)[c].size();
     }
 
     *num_nmsed_out = num_det;
     const T* scores_data = scores.data<T>();
     if (keep_top_k > -1 && num_det > keep_top_k) {
+      const T* sdata;
       std::vector<std::pair<float, std::pair<int, int>>> score_index_pairs;
       for (const auto& it : *indices) {
         int label = it.first;
-        const T* sdata = scores_data + label * predict_dim;
+        if (scores_size == 3) {
+          sdata = scores_data + label * scores.dims()[1];
+        } else {
+          score_slice.Resize({scores.dims()[0], 1});
+          SliceOneClass<T>(dev_ctx, scores, label, &score_slice);
+          sdata = score_slice.data<T>();
+        }
         const std::vector<int>& label_indices = it.second;
         for (size_t j = 0; j < label_indices.size(); ++j) {
           int idx = label_indices[j];
-          PADDLE_ENFORCE_LT(idx, predict_dim);
           score_index_pairs.push_back(
               std::make_pair(sdata[idx], std::make_pair(label, idx)));
         }
@@ -252,31 +313,55 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
         int idx = score_index_pairs[j].second.second;
         new_indices[label].push_back(idx);
       }
+      if (scores_size == 2) {
+        for (const auto& it : new_indices) {
+          int label = it.first;
+          std::stable_sort(new_indices[label].begin(),
+                           new_indices[label].end());
+        }
+      }
       new_indices.swap(*indices);
       *num_nmsed_out = keep_top_k;
     }
   }
 
-  void MultiClassOutput(const Tensor& scores, const Tensor& bboxes,
+  void MultiClassOutput(const platform::DeviceContext& ctx,
+                        const Tensor& scores, const Tensor& bboxes,
                         const std::map<int, std::vector<int>>& selected_indices,
-                        Tensor* outs) const {
+                        const int scores_size, Tensor* outs) const {
+    int64_t class_num = scores.dims()[1];
     int64_t predict_dim = scores.dims()[1];
     int64_t box_size = bboxes.dims()[1];
-    int64_t out_dim = bboxes.dims()[1] + 2;
+    if (scores_size == 2) {
+      box_size = bboxes.dims()[2];
+    }
+    int64_t out_dim = box_size + 2;
     auto* scores_data = scores.data<T>();
     auto* bboxes_data = bboxes.data<T>();
     auto* odata = outs->data<T>();
-
+    const T* sdata;
+    Tensor bbox;
+    bbox.Resize({scores.dims()[0], box_size});
     int count = 0;
     for (const auto& it : selected_indices) {
       int label = it.first;
-      const T* sdata = scores_data + label * predict_dim;
       const std::vector<int>& indices = it.second;
+      if (scores_size == 2) {
+        SliceOneClass<T>(ctx, bboxes, label, &bbox);
+      } else {
+        sdata = scores_data + label * predict_dim;
+      }
       for (size_t j = 0; j < indices.size(); ++j) {
         int idx = indices[j];
-        const T* bdata = bboxes_data + idx * box_size;
-        odata[count * out_dim] = label;           // label
-        odata[count * out_dim + 1] = sdata[idx];  // score
+        odata[count * out_dim] = label;  // label
+        const T* bdata;
+        if (scores_size == 3) {
+          bdata = bboxes_data + idx * box_size;
+          odata[count * out_dim + 1] = sdata[idx];  // score
+        } else {
+          bdata = bbox.data<T>() + idx * box_size;
+          odata[count * out_dim + 1] = *(scores_data + idx * class_num + label);
+        }
         // xmin, ymin, xmax, ymax or multi-points coordinates
         std::memcpy(odata + count * out_dim + 2, bdata, box_size * sizeof(T));
         count++;
@@ -285,52 +370,64 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
   }
 
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* boxes = ctx.Input<Tensor>("BBoxes");
-    auto* scores = ctx.Input<Tensor>("Scores");
+    auto* boxes = ctx.Input<LoDTensor>("BBoxes");
+    auto* scores = ctx.Input<LoDTensor>("Scores");
     auto* outs = ctx.Output<LoDTensor>("Out");
 
     auto score_dims = scores->dims();
-
-    int64_t batch_size = score_dims[0];
-    int64_t class_num = score_dims[1];
-    int64_t predict_dim = score_dims[2];
-    int64_t box_dim = boxes->dims()[2];
-    int64_t out_dim = boxes->dims()[2] + 2;
+    auto score_size = score_dims.size();
+    auto& dev_ctx = ctx.template device_context<platform::CPUDeviceContext>();
 
     std::vector<std::map<int, std::vector<int>>> all_indices;
     std::vector<size_t> batch_starts = {0};
-    for (int64_t i = 0; i < batch_size; ++i) {
-      Tensor ins_score = scores->Slice(i, i + 1);
-      ins_score.Resize({class_num, predict_dim});
-
-      Tensor ins_boxes = boxes->Slice(i, i + 1);
-      ins_boxes.Resize({predict_dim, box_dim});
-
+    int64_t batch_size = score_dims[0];
+    int64_t box_dim = boxes->dims()[2];
+    int64_t out_dim = box_dim + 2;
+    int num_nmsed_out = 0;
+    Tensor boxes_slice, scores_slice;
+    int n = score_size == 3 ? batch_size : boxes->lod().back().size() - 1;
+    for (int i = 0; i < n; ++i) {
+      if (score_size == 3) {
+        scores_slice = scores->Slice(i, i + 1);
+        scores_slice.Resize({score_dims[1], score_dims[2]});
+        boxes_slice = boxes->Slice(i, i + 1);
+        boxes_slice.Resize({score_dims[2], box_dim});
+      } else {
+        auto boxes_lod = boxes->lod().back();
+        scores_slice = scores->Slice(boxes_lod[i], boxes_lod[i + 1]);
+        boxes_slice = boxes->Slice(boxes_lod[i], boxes_lod[i + 1]);
+      }
       std::map<int, std::vector<int>> indices;
-      int num_nmsed_out = 0;
-      MultiClassNMS(ctx, ins_score, ins_boxes, &indices, &num_nmsed_out);
+      MultiClassNMS(ctx, scores_slice, boxes_slice, score_size, &indices,
+                    &num_nmsed_out);
       all_indices.push_back(indices);
       batch_starts.push_back(batch_starts.back() + num_nmsed_out);
     }
 
     int num_kept = batch_starts.back();
     if (num_kept == 0) {
-      T* od = outs->mutable_data<T>({1}, ctx.GetPlace());
+      T* od = outs->mutable_data<T>({1, 1}, ctx.GetPlace());
       od[0] = -1;
+      batch_starts = {0, 1};
     } else {
       outs->mutable_data<T>({num_kept, out_dim}, ctx.GetPlace());
-      for (int64_t i = 0; i < batch_size; ++i) {
-        Tensor ins_score = scores->Slice(i, i + 1);
-        ins_score.Resize({class_num, predict_dim});
-
-        Tensor ins_boxes = boxes->Slice(i, i + 1);
-        ins_boxes.Resize({predict_dim, box_dim});
-
+      for (int i = 0; i < n; ++i) {
+        if (score_size == 3) {
+          scores_slice = scores->Slice(i, i + 1);
+          boxes_slice = boxes->Slice(i, i + 1);
+          scores_slice.Resize({score_dims[1], score_dims[2]});
+          boxes_slice.Resize({score_dims[2], box_dim});
+        } else {
+          auto boxes_lod = boxes->lod().back();
+          scores_slice = scores->Slice(boxes_lod[i], boxes_lod[i + 1]);
+          boxes_slice = boxes->Slice(boxes_lod[i], boxes_lod[i + 1]);
+        }
         int64_t s = batch_starts[i];
         int64_t e = batch_starts[i + 1];
         if (e > s) {
           Tensor out = outs->Slice(s, e);
-          MultiClassOutput(ins_score, ins_boxes, all_indices[i], &out);
+          MultiClassOutput(dev_ctx, scores_slice, boxes_slice, all_indices[i],
+                           score_dims.size(), &out);
         }
       }
     }
@@ -346,17 +443,24 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
     AddInput("BBoxes",
-             "(Tensor) A 3-D Tensor with shape "
+             "Two types of bboxes are supported:"
+             "1. (Tensor) A 3-D Tensor with shape "
              "[N, M, 4 or 8 16 24 32] represents the "
              "predicted locations of M bounding bboxes, N is the batch size. "
              "Each bounding box has four coordinate values and the layout is "
-             "[xmin, ymin, xmax, ymax], when box size equals to 4.");
+             "[xmin, ymin, xmax, ymax], when box size equals to 4."
+             "2. (LoDTensor) A 3-D Tensor with shape [M, C, 4]"
+             "M is the number of bounding boxes, C is the class number");
     AddInput("Scores",
-             "(Tensor) A 3-D Tensor with shape [N, C, M] represents the "
+             "Two types of scores are supported:"
+             "1. (Tensor) A 3-D Tensor with shape [N, C, M] represents the "
              "predicted confidence predictions. N is the batch size, C is the "
              "class number, M is number of bounding boxes. For each category "
              "there are total M scores which corresponding M bounding boxes. "
-             " Please note, M is equal to the 1st dimension of BBoxes. ");
+             " Please note, M is equal to the 2nd dimension of BBoxes. "
+             "2. (LoDTensor) A 2-D LoDTensor with shape [M, C]. "
+             "M is the number of bbox, C is the class number. In this case, "
+             "Input BBoxes should be the second case with shape [M, C, 4].");
     AddAttr<int>(
         "background_label",
         "(int, defalut: 0) "
@@ -384,6 +488,10 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker {
                  "(int64_t) "
                  "Number of total bboxes to be kept per image after NMS "
                  "step. -1 means keeping all bboxes after NMS step.");
+    AddAttr<bool>("normalized",
+                  "(bool, default true) "
+                  "Whether detections are normalized.")
+        .SetDefault(true);
     AddOutput("Out",
               "(LoDTensor) A 2-D LoDTensor with shape [No, 6] represents the "
               "detections. Each row has 6 values: "
@@ -399,24 +507,21 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(
 This operator is to do multi-class non maximum suppression (NMS) on a batched
 of boxes and scores.
-
 In the NMS step, this operator greedily selects a subset of detection bounding
 boxes that have high scores larger than score_threshold, if providing this
 threshold, then selects the largest nms_top_k confidences scores if nms_top_k
 is larger than -1. Then this operator pruns away boxes that have high IOU
 (intersection over union) overlap with already selected boxes by adaptive
 threshold NMS based on parameters of nms_threshold and nms_eta.
-
 Aftern NMS step, at most keep_top_k number of total bboxes are to be kept
 per image if keep_top_k is larger than -1.
-
 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. If there is no detected boxes
-for all images, all the elements in LoD are 0, and the Out only contains one
-value which is -1.
+for all images, all the elements in LoD are set to {1}, and the Out only 
+contains one value which is -1.
 )DOC");
   }
 };

python/paddle/fluid/layers/detection.py

@@ -49,6 +49,7 @@ __all__ = [
     'box_coder',
     'polygon_box_transform',
     'yolov3_loss',
+    'multiclass_nms',
 ]
 
 
@@ -262,8 +263,10 @@ def detection_output(loc,
             number is N + 1, N is the batch size. The i-th image has
             `LoD[i + 1] - LoD[i]` detected results, if it is 0, the i-th image
             has no detected results. If all images have not detected results,
-            all the elements in LoD are 0, and output tensor only contains one
+            LoD will be set to {1}, and output tensor only contains one
             value, which is -1.
+            (After version 1.3, when no boxes detected, the lod is changed
+             from {0} to {1}.)
 
     Examples:
         .. code-block:: python
@@ -1960,3 +1963,119 @@ def generate_proposals(scores,
     rpn_roi_probs.stop_gradient = True
 
     return rpn_rois, rpn_roi_probs
+
+
+def multiclass_nms(bboxes,
+                   scores,
+                   score_threshold,
+                   nms_top_k,
+                   keep_top_k,
+                   nms_threshold=0.3,
+                   normalized=True,
+                   nms_eta=1.,
+                   background_label=0,
+                   name=None):
+    """
+    **Multiclass NMS**
+    
+    This operator is to do multi-class non maximum suppression (NMS) on
+    boxes and scores.
+
+    In the NMS step, this operator greedily selects a subset of detection bounding
+    boxes that have high scores larger than score_threshold, if providing this
+    threshold, then selects the largest nms_top_k confidences scores if nms_top_k
+    is larger than -1. Then this operator pruns away boxes that have high IOU
+    (intersection over union) overlap with already selected boxes by adaptive
+    threshold NMS based on parameters of nms_threshold and nms_eta.
+
+    Aftern NMS step, at most keep_top_k number of total bboxes are to be kept
+    per image if keep_top_k is larger than -1.
+
+    Args:
+        bboxes (Variable): Two types of bboxes are supported:
+                           1. (Tensor) A 3-D Tensor with shape
+                           [N, M, 4 or 8 16 24 32] represents the
+                           predicted locations of M bounding bboxes,
+                           N is the batch size. Each bounding box has four
+                           coordinate values and the layout is 
+                           [xmin, ymin, xmax, ymax], when box size equals to 4.
+                           2. (LoDTensor) A 3-D Tensor with shape [M, C, 4]
+                           M is the number of bounding boxes, C is the 
+                           class number   
+        scores (Variable): Two types of scores are supported:
+                           1. (Tensor) A 3-D Tensor with shape [N, C, M]
+                           represents the predicted confidence predictions.
+                           N is the batch size, C is the class number, M is 
+                           number of bounding boxes. For each category there 
+                           are total M scores which corresponding M bounding
+                           boxes. Please note, M is equal to the 2nd dimension
+                           of BBoxes.
+                           2. (LoDTensor) A 2-D LoDTensor with shape [M, C].
+                           M is the number of bbox, C is the class number.
+                           In this case, input BBoxes should be the second
+                           case with shape [M, C, 4].
+        background_label (int): The index of background label, the background 
+                                label will be ignored. If set to -1, then all
+                                categories will be considered. Default: 0
+        score_threshold (float): Threshold to filter out bounding boxes with
+                                 low confidence score. If not provided, 
+                                 consider all boxes.
+        nms_top_k (int): Maximum number of detections to be kept according to
+                         the confidences aftern the filtering detections based
+                         on score_threshold.
+        nms_threshold (float): The threshold to be used in NMS. Default: 0.3
+        nms_eta (float): The threshold to be used in NMS. Default: 1.0
+        keep_top_k (int): Number of total bboxes to be kept per image after NMS
+                          step. -1 means keeping all bboxes after NMS step.
+        normalized (bool): Whether detections are normalized. Default: True
+        name(str): Name of the multiclass nms op. Default: None.
+
+    Returns:
+        Out: A 2-D LoDTensor with shape [No, 6] represents the detections.
+             Each row has 6 values: [label, confidence, xmin, ymin, xmax, ymax]
+             or A 2-D LoDTensor with shape [No, 10] represents the detections.
+             Each row has 10 values: 
+             [label, confidence, x1, y1, x2, y2, x3, y3, x4, y4]. No is the 
+             total number of detections. If there is no detected boxes for all
+             images, lod will be set to {1} and Out only contains one value
+             which is -1.
+             (After version 1.3, when no boxes detected, the lod is changed 
+             from {0} to {1}) 
+
+    Examples:
+        .. code-block:: python
+
+            boxes = fluid.layers.data(name='bboxes', shape=[81, 4],
+                                      dtype='float32', lod_level=1)
+            scores = fluid.layers.data(name='scores', shape=[81],
+                                      dtype='float32', lod_level=1)
+            out = fluid.layers.multiclass_nms(bboxes=boxes,
+                                              scores=scores,
+                                              background_label=0,
+                                              score_threshold=0.5,
+                                              nms_top_k=400,
+                                              nms_threshold=0.3,
+                                              keep_top_k=200,
+                                              normalized=False)
+    """
+    helper = LayerHelper('multiclass_nms', **locals())
+
+    output = helper.create_variable_for_type_inference(dtype=bboxes.dtype)
+    helper.append_op(
+        type="multiclass_nms",
+        inputs={'BBoxes': bboxes,
+                'Scores': scores},
+        attrs={
+            'background_label': background_label,
+            'score_threshold': score_threshold,
+            'nms_top_k': nms_top_k,
+            'nms_threshold': nms_threshold,
+            'nms_eta': nms_eta,
+            'keep_top_k': keep_top_k,
+            'nms_eta': nms_eta,
+            'normalized': normalized
+        },
+        outputs={'Out': output})
+    output.stop_gradient = True
+
+    return output

python/paddle/fluid/tests/test_detection.py

@@ -469,5 +469,16 @@ class TestYoloDetection(unittest.TestCase):
             self.assertIsNotNone(loss)
 
 
+class TestMulticlassNMS(unittest.TestCase):
+    def test_multiclass_nms(self):
+        program = Program()
+        with program_guard(program):
+            bboxes = layers.data(
+                name='bboxes', shape=[-1, 10, 4], dtype='float32')
+            scores = layers.data(name='scores', shape=[-1, 10], dtype='float32')
+            output = layers.multiclass_nms(bboxes, scores, 0.3, 400, 200, 0.7)
+            self.assertIsNotNone(output)
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py

@@ -19,7 +19,7 @@ import copy
 from op_test import OpTest
 
 
-def iou(box_a, box_b):
+def iou(box_a, box_b, norm):
     """Apply intersection-over-union overlap between box_a and box_b
     """
     xmin_a = min(box_a[0], box_a[2])
@@ -32,8 +32,10 @@ def iou(box_a, box_b):
     xmax_b = max(box_b[0], box_b[2])
     ymax_b = max(box_b[1], box_b[3])
 
-    area_a = (ymax_a - ymin_a) * (xmax_a - xmin_a)
-    area_b = (ymax_b - ymin_b) * (xmax_b - xmin_b)
+    area_a = (ymax_a - ymin_a + (norm == False)) * (xmax_a - xmin_a +
+                                                    (norm == False))
+    area_b = (ymax_b - ymin_b + (norm == False)) * (xmax_b - xmin_b +
+                                                    (norm == False))
     if area_a <= 0 and area_b <= 0:
         return 0.0
 
@@ -42,17 +44,21 @@ def iou(box_a, box_b):
     xb = min(xmax_a, xmax_b)
     yb = min(ymax_a, ymax_b)
 
-    inter_area = max(xb - xa, 0.0) * max(yb - ya, 0.0)
-
-    box_a_area = (box_a[2] - box_a[0]) * (box_a[3] - box_a[1])
-    box_b_area = (box_b[2] - box_b[0]) * (box_b[3] - box_b[1])
+    inter_area = max(xb - xa + (norm == False),
+                     0.0) * max(yb - ya + (norm == False), 0.0)
 
     iou_ratio = inter_area / (area_a + area_b - inter_area)
 
     return iou_ratio
 
 
-def nms(boxes, scores, score_threshold, nms_threshold, top_k=200, eta=1.0):
+def nms(boxes,
+        scores,
+        score_threshold,
+        nms_threshold,
+        top_k=200,
+        normalized=True,
+        eta=1.0):
     """Apply non-maximum suppression at test time to avoid detecting too many
     overlapping bounding boxes for a given object.
     Args:
@@ -87,7 +93,7 @@ def nms(boxes, scores, score_threshold, nms_threshold, top_k=200, eta=1.0):
         for k in range(len(selected_indices)):
             if keep:
                 kept_idx = selected_indices[k]
-                overlap = iou(boxes[idx], boxes[kept_idx])
+                overlap = iou(boxes[idx], boxes[kept_idx], normalized)
                 keep = True if overlap <= adaptive_threshold else False
             else:
                 break
@@ -99,16 +105,24 @@ def nms(boxes, scores, score_threshold, nms_threshold, top_k=200, eta=1.0):
 
 
 def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold,
-                   nms_top_k, keep_top_k):
-    class_num = scores.shape[0]
-    priorbox_num = scores.shape[1]
+                   nms_top_k, keep_top_k, normalized, shared):
+    if shared:
+        class_num = scores.shape[0]
+        priorbox_num = scores.shape[1]
+    else:
+        box_num = scores.shape[0]
+        class_num = scores.shape[1]
 
     selected_indices = {}
     num_det = 0
     for c in range(class_num):
         if c == background: continue
-        indices = nms(boxes, scores[c], score_threshold, nms_threshold,
-                      nms_top_k)
+        if shared:
+            indices = nms(boxes, scores[c], score_threshold, nms_threshold,
+                          nms_top_k, normalized)
+        else:
+            indices = nms(boxes[:, c, :], scores[:, c], score_threshold,
+                          nms_threshold, nms_top_k, normalized)
         selected_indices[c] = indices
         num_det += len(indices)
 
@@ -116,7 +130,10 @@ def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold,
         score_index = []
         for c, indices in selected_indices.items():
             for idx in indices:
-                score_index.append((scores[c][idx], c, idx))
+                if shared:
+                    score_index.append((scores[c][idx], c, idx))
+                else:
+                    score_index.append((scores[idx][c], c, idx))
 
         sorted_score_index = sorted(
             score_index, key=lambda tup: tup[0], reverse=True)
@@ -127,24 +144,75 @@ def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold,
             selected_indices[c] = []
         for s, c, idx in sorted_score_index:
             selected_indices[c].append(idx)
+        if not shared:
+            for labels in selected_indices:
+                selected_indices[labels].sort()
         num_det = keep_top_k
 
     return selected_indices, num_det
 
 
-def batched_multiclass_nms(boxes, scores, background, score_threshold,
-                           nms_threshold, nms_top_k, keep_top_k):
+def lod_multiclass_nms(boxes, scores, background, score_threshold,
+                       nms_threshold, nms_top_k, keep_top_k, box_lod,
+                       normalized):
+    det_outs = []
+    lod = []
+    head = 0
+    for n in range(len(box_lod[0])):
+        box = boxes[head:head + box_lod[0][n]]
+        score = scores[head:head + box_lod[0][n]]
+        head = head + box_lod[0][n]
+        nmsed_outs, nmsed_num = multiclass_nms(
+            box,
+            score,
+            background,
+            score_threshold,
+            nms_threshold,
+            nms_top_k,
+            keep_top_k,
+            normalized,
+            shared=False)
+        if nmsed_num == 0:
+            #lod.append(1)
+            continue
+        lod.append(nmsed_num)
+        for c, indices in nmsed_outs.items():
+            for idx in indices:
+                xmin, ymin, xmax, ymax = box[idx, c, :]
+                det_outs.append([c, score[idx][c], xmin, ymin, xmax, ymax])
+    if len(lod) == 0:
+        lod.append(1)
+
+    return det_outs, lod
+
+
+def batched_multiclass_nms(boxes,
+                           scores,
+                           background,
+                           score_threshold,
+                           nms_threshold,
+                           nms_top_k,
+                           keep_top_k,
+                           normalized=True):
     batch_size = scores.shape[0]
 
     det_outs = []
     lod = []
     for n in range(batch_size):
-        nmsed_outs, nmsed_num = multiclass_nms(boxes[n], scores[n], background,
-                                               score_threshold, nms_threshold,
-                                               nms_top_k, keep_top_k)
-        lod.append(nmsed_num)
-        if nmsed_num == 0: continue
+        nmsed_outs, nmsed_num = multiclass_nms(
+            boxes[n],
+            scores[n],
+            background,
+            score_threshold,
+            nms_threshold,
+            nms_top_k,
+            keep_top_k,
+            normalized,
+            shared=True)
+        if nmsed_num == 0:
+            continue
 
+        lod.append(nmsed_num)
         tmp_det_out = []
         for c, indices in nmsed_outs.items():
             for idx in indices:
@@ -154,7 +222,8 @@ def batched_multiclass_nms(boxes, scores, background, score_threshold,
         sorted_det_out = sorted(
             tmp_det_out, key=lambda tup: tup[0], reverse=False)
         det_outs.extend(sorted_det_out)
-
+    if len(lod) == 0:
+        lod += [1]
     return det_outs, lod
 
 
@@ -168,7 +237,6 @@ class TestMulticlassNMSOp(OpTest):
         M = 1200
         C = 21
         BOX_SIZE = 4
-
         background = 0
         nms_threshold = 0.3
         nms_top_k = 400
@@ -206,6 +274,7 @@ class TestMulticlassNMSOp(OpTest):
             'keep_top_k': keep_top_k,
             'score_threshold': score_threshold,
             'nms_eta': 1.0,
+            'normalized': True,
         }
 
     def test_check_output(self):
@@ -219,13 +288,70 @@ class TestMulticlassNMSOpNoOutput(TestMulticlassNMSOp):
         self.score_threshold = 2.0
 
 
+class TestMulticlassNMSLoDInput(OpTest):
+    def set_argument(self):
+        self.score_threshold = 0.01
+
+    def setUp(self):
+        self.set_argument()
+        M = 1200
+        C = 21
+        BOX_SIZE = 4
+        box_lod = [[1200]]
+        background = 0
+        nms_threshold = 0.3
+        nms_top_k = 400
+        keep_top_k = 200
+        score_threshold = self.score_threshold
+        normalized = False
+
+        scores = np.random.random((M, C)).astype('float32')
+
+        def softmax(x):
+            shiftx = x - np.max(x).clip(-64.)
+            exps = np.exp(shiftx)
+            return exps / np.sum(exps)
+
+        scores = np.apply_along_axis(softmax, 1, scores)
+
+        boxes = np.random.random((M, C, BOX_SIZE)).astype('float32')
+        boxes[:, :, 0] = boxes[:, :, 0] * 10
+        boxes[:, :, 1] = boxes[:, :, 1] * 10
+        boxes[:, :, 2] = boxes[:, :, 2] * 10 + 10
+        boxes[:, :, 3] = boxes[:, :, 3] * 10 + 10
+
+        nmsed_outs, lod = lod_multiclass_nms(
+            boxes, scores, background, score_threshold, nms_threshold,
+            nms_top_k, keep_top_k, box_lod, normalized)
+        nmsed_outs = [-1] if not nmsed_outs else nmsed_outs
+        nmsed_outs = np.array(nmsed_outs).astype('float32')
+        self.op_type = 'multiclass_nms'
+        self.inputs = {
+            'BBoxes': (boxes, box_lod),
+            'Scores': (scores, box_lod),
+        }
+        self.outputs = {'Out': (nmsed_outs, [lod])}
+        self.attrs = {
+            'background_label': 0,
+            'nms_threshold': nms_threshold,
+            'nms_top_k': nms_top_k,
+            'keep_top_k': keep_top_k,
+            'score_threshold': score_threshold,
+            'nms_eta': 1.0,
+            'normalized': normalized,
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
 class TestIOU(unittest.TestCase):
     def test_iou(self):
         box1 = np.array([4.0, 3.0, 7.0, 5.0]).astype('float32')
         box2 = np.array([3.0, 4.0, 6.0, 8.0]).astype('float32')
 
         expt_output = np.array([2.0 / 16.0]).astype('float32')
-        calc_output = np.array([iou(box1, box2)]).astype('float32')
+        calc_output = np.array([iou(box1, box2, True)]).astype('float32')
         self.assertTrue(np.allclose(calc_output, expt_output))