update code from review comments, test=develop

Author: jerrywgz

paddle/fluid/operators/detection/bbox_util.h

@@ -149,17 +149,18 @@ void ClipTiledBoxes(const platform::DeviceContext& ctx,
   }
 }
 
+// Calculate max IoU between each box and ground-truth and
+// each row represents one box
 template <typename T>
-void MaxOverlaps(const framework::Tensor& iou,
-                 framework::Tensor* max_overlaps) {
-  const T* proposal_to_gt_overlaps = iou.data<T>();
+void MaxIoU(const framework::Tensor& iou, framework::Tensor* max_iou) {
+  const T* iou_data = iou.data<T>();
   int row = iou.dims()[0];
   int col = iou.dims()[1];
-  T* max_overlaps_dt = max_overlaps->data<T>();
+  T* max_iou_data = max_iou->data<T>();
   for (int i = 0; i < row; ++i) {
-    const T* v = proposal_to_gt_overlaps + i * col;
+    const T* v = iou_data + i * col;
     T max_v = *std::max_element(v, v + col);
-    max_overlaps_dt[i] = max_v;
+    max_iou_data[i] = max_v;
   }
 }
 

paddle/fluid/operators/detection/generate_proposal_labels_op.cc

@@ -33,6 +33,9 @@ void AppendRois(LoDTensor* out, int64_t offset, Tensor* to_add) {
   memcpy(out_data + offset, to_add_data, to_add->numel() * sizeof(T));
 }
 
+// Filter the ground-truth in RoIs and the RoIs with non-positive area.
+// The ground-truth has max overlap with itself so the max_overlap is 1
+// and the corresponding RoI will be removed.
 template <typename T>
 void FilterRoIs(const platform::DeviceContext& ctx, const Tensor& rpn_rois,
                 const Tensor& max_overlap, Tensor* keep) {
@@ -354,7 +357,7 @@ std::vector<Tensor> SampleRoisForOneImage(
   proposal_with_max_overlap.mutable_data<T>({proposals_num},
                                             context.GetPlace());
 
-  MaxOverlaps<T>(proposal_to_gt_overlaps, &proposal_with_max_overlap);
+  MaxIoU<T>(proposal_to_gt_overlaps, &proposal_with_max_overlap);
 
   // Generate proposal index
   std::vector<std::vector<int>> fg_bg_gt =
@@ -546,7 +549,7 @@ class GenerateProposalLabelsKernel : public framework::OpKernel<T> {
         max_overlap_slice =
             max_overlap->Slice(rpn_rois_lod[i], rpn_rois_lod[i + 1]);
       } else {
-        max_overlap_slice.mutable_data<T>(im_info_slice.dims(),
+        max_overlap_slice.mutable_data<T>({rpn_rois_slice.dims()[0]},
                                           context.GetPlace());
       }
       std::vector<Tensor> tensor_output = SampleRoisForOneImage<T>(
@@ -620,8 +623,8 @@ class GenerateProposalLabelsOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("MaxOverlap",
              "(LoDTensor), This input is a 1D LoDTensor with shape [N]."
              "N is the number of Input(RpnRois), "
-             "each element is the maxoverlap between "
-             "the input RoIs and ground-truth.")
+             "each element is the maximum overlap between "
+             "the proposal RoI and ground-truth.")
         .AsDispensable();
 
     AddOutput(

python/paddle/fluid/layers/detection.py

@@ -2640,8 +2640,8 @@ def generate_proposal_labels(rpn_rois,
         use_random(bool): Use random sampling to choose foreground and background boxes.
         is_cls_agnostic(bool): bbox regression use class agnostic simply which only represent fg and bg boxes.
         is_cascade_rcnn(bool): it will filter some bbox crossing the image's boundary when setting True.
-        max_overlap(Variable): Maximum overlap between each Input box and ground-truth.
-        return_max_overlap(bool): Whether return the maximum overlap between each Output box and ground-truth.
+        max_overlap(Variable): Maximum overlap between each proposal box and ground-truth.
+        return_max_overlap(bool): Whether return the maximum overlap between each sampled RoI and ground-truth.
 
     Returns:
         tuple:
@@ -2652,7 +2652,7 @@ def generate_proposal_labels(rpn_rois,
         - **bbox_targets**: 2-D LoDTensor with shape ``[batch_size_per_im * batch_size, 4 * class_num]``. The regression targets of all RoIs. The data type is the same as ``rpn_rois``.
         - **bbox_inside_weights**: 2-D LoDTensor with shape ``[batch_size_per_im * batch_size, 4 * class_num]``. The weights of foreground boxes' regression loss. The data type is the same as ``rpn_rois``.
         - **bbox_outside_weights**: 2-D LoDTensor with shape ``[batch_size_per_im * batch_size, 4 * class_num]``. The weights of regression loss. The data type is the same as ``rpn_rois``.
-        - **max_overlap**: 1-D LoDTensor with shape ``[P]``. P is the number of output ``rois``. The maximum overlap between each output box and ground-truth.
+        - **max_overlap**: 1-D LoDTensor with shape ``[P]``. P is the number of output ``rois``. The maximum overlap between each sampled RoI and ground-truth.
 
     Examples:
         .. code-block:: python

python/paddle/fluid/tests/test_detection.py

@@ -318,8 +318,7 @@ def test_generate_proposal_labels(self):
                 name='is_crowd', shape=[6], dtype='int32', lod_level=1)
             gt_boxes = fluid.data(
                 name='gt_boxes', shape=[6, 4], dtype='float32', lod_level=1)
-            im_info = fluid.data(
-                name='im_info', shape=[1, 3], dtype='float32', lod_level=1)
+            im_info = fluid.data(name='im_info', shape=[1, 3], dtype='float32')
             max_overlap = fluid.data(
                 name='max_overlap', shape=[4], dtype='float32', lod_level=1)
             self.class_nums = 5