Merge pull request #10255 from dkurt:dnn_roi_pooling

modules/dnn/include/opencv2/dnn/all_layers.hpp

@@ -242,7 +242,8 @@ CV__DNN_EXPERIMENTAL_NS_BEGIN
         {
             MAX,
             AVE,
-            STOCHASTIC
+            STOCHASTIC,
+            ROI
         };
 
         int type;
@@ -251,6 +252,9 @@ CV__DNN_EXPERIMENTAL_NS_BEGIN
         bool computeMaxIdx;
         String padMode;
         bool ceilMode;
+        // ROIPooling parameters.
+        Size pooledSize;
+        float spatialScale;
 
         static Ptr<PoolingLayer> create(const LayerParams& params);
     };

modules/dnn/misc/caffe/opencv-caffe.pb.h

@@ -87,6 +87,7 @@ class PoolingParameter;
 class PowerParameter;
 class PriorBoxParameter;
 class PythonParameter;
+class ROIPoolingParameter;
 class ReLUParameter;
 class RecurrentParameter;
 class ReductionParameter;
@@ -4182,6 +4183,15 @@ class LayerParameter : public ::google::protobuf::Message /* @@protoc_insertion_
   ::opencv_caffe::ReshapeParameter* release_reshape_param();
   void set_allocated_reshape_param(::opencv_caffe::ReshapeParameter* reshape_param);
 
+  // optional .opencv_caffe.ROIPoolingParameter roi_pooling_param = 8266711;
+  bool has_roi_pooling_param() const;
+  void clear_roi_pooling_param();
+  static const int kRoiPoolingParamFieldNumber = 8266711;
+  const ::opencv_caffe::ROIPoolingParameter& roi_pooling_param() const;
+  ::opencv_caffe::ROIPoolingParameter* mutable_roi_pooling_param();
+  ::opencv_caffe::ROIPoolingParameter* release_roi_pooling_param();
+  void set_allocated_roi_pooling_param(::opencv_caffe::ROIPoolingParameter* roi_pooling_param);
+
   // optional .opencv_caffe.ScaleParameter scale_param = 142;
   bool has_scale_param() const;
   void clear_scale_param();
@@ -4355,6 +4365,8 @@ class LayerParameter : public ::google::protobuf::Message /* @@protoc_insertion_
   inline void clear_has_relu_param();
   inline void set_has_reshape_param();
   inline void clear_has_reshape_param();
+  inline void set_has_roi_pooling_param();
+  inline void clear_has_roi_pooling_param();
   inline void set_has_scale_param();
   inline void clear_has_scale_param();
   inline void set_has_sigmoid_param();
@@ -4428,6 +4440,7 @@ class LayerParameter : public ::google::protobuf::Message /* @@protoc_insertion_
   ::opencv_caffe::ReductionParameter* reduction_param_;
   ::opencv_caffe::ReLUParameter* relu_param_;
   ::opencv_caffe::ReshapeParameter* reshape_param_;
+  ::opencv_caffe::ROIPoolingParameter* roi_pooling_param_;
   ::opencv_caffe::ScaleParameter* scale_param_;
   ::opencv_caffe::SigmoidParameter* sigmoid_param_;
   ::opencv_caffe::SoftmaxParameter* softmax_param_;
@@ -12783,6 +12796,124 @@ class NormalizedBBox : public ::google::protobuf::Message /* @@protoc_insertion_
 };
 extern ::google::protobuf::internal::ExplicitlyConstructed<NormalizedBBox> NormalizedBBox_default_instance_;
 
+// -------------------------------------------------------------------
+
+class ROIPoolingParameter : public ::google::protobuf::Message /* @@protoc_insertion_point(class_definition:opencv_caffe.ROIPoolingParameter) */ {
+ public:
+  ROIPoolingParameter();
+  virtual ~ROIPoolingParameter();
+
+  ROIPoolingParameter(const ROIPoolingParameter& from);
+
+  inline ROIPoolingParameter& operator=(const ROIPoolingParameter& from) {
+    CopyFrom(from);
+    return *this;
+  }
+
+  inline const ::google::protobuf::UnknownFieldSet& unknown_fields() const {
+    return _internal_metadata_.unknown_fields();
+  }
+
+  inline ::google::protobuf::UnknownFieldSet* mutable_unknown_fields() {
+    return _internal_metadata_.mutable_unknown_fields();
+  }
+
+  static const ::google::protobuf::Descriptor* descriptor();
+  static const ROIPoolingParameter& default_instance();
+
+  static const ROIPoolingParameter* internal_default_instance();
+
+  void Swap(ROIPoolingParameter* other);
+
+  // implements Message ----------------------------------------------
+
+  inline ROIPoolingParameter* New() const { return New(NULL); }
+
+  ROIPoolingParameter* New(::google::protobuf::Arena* arena) const;
+  void CopyFrom(const ::google::protobuf::Message& from);
+  void MergeFrom(const ::google::protobuf::Message& from);
+  void CopyFrom(const ROIPoolingParameter& from);
+  void MergeFrom(const ROIPoolingParameter& from);
+  void Clear();
+  bool IsInitialized() const;
+
+  size_t ByteSizeLong() const;
+  bool MergePartialFromCodedStream(
+      ::google::protobuf::io::CodedInputStream* input);
+  void SerializeWithCachedSizes(
+      ::google::protobuf::io::CodedOutputStream* output) const;
+  ::google::protobuf::uint8* InternalSerializeWithCachedSizesToArray(
+      bool deterministic, ::google::protobuf::uint8* output) const;
+  ::google::protobuf::uint8* SerializeWithCachedSizesToArray(::google::protobuf::uint8* output) const {
+    return InternalSerializeWithCachedSizesToArray(false, output);
+  }
+  int GetCachedSize() const { return _cached_size_; }
+  private:
+  void SharedCtor();
+  void SharedDtor();
+  void SetCachedSize(int size) const;
+  void InternalSwap(ROIPoolingParameter* other);
+  void UnsafeMergeFrom(const ROIPoolingParameter& from);
+  private:
+  inline ::google::protobuf::Arena* GetArenaNoVirtual() const {
+    return _internal_metadata_.arena();
+  }
+  inline void* MaybeArenaPtr() const {
+    return _internal_metadata_.raw_arena_ptr();
+  }
+  public:
+
+  ::google::protobuf::Metadata GetMetadata() const;
+
+  // nested types ----------------------------------------------------
+
+  // accessors -------------------------------------------------------
+
+  // optional uint32 pooled_h = 1 [default = 0];
+  bool has_pooled_h() const;
+  void clear_pooled_h();
+  static const int kPooledHFieldNumber = 1;
+  ::google::protobuf::uint32 pooled_h() const;
+  void set_pooled_h(::google::protobuf::uint32 value);
+
+  // optional uint32 pooled_w = 2 [default = 0];
+  bool has_pooled_w() const;
+  void clear_pooled_w();
+  static const int kPooledWFieldNumber = 2;
+  ::google::protobuf::uint32 pooled_w() const;
+  void set_pooled_w(::google::protobuf::uint32 value);
+
+  // optional float spatial_scale = 3 [default = 1];
+  bool has_spatial_scale() const;
+  void clear_spatial_scale();
+  static const int kSpatialScaleFieldNumber = 3;
+  float spatial_scale() const;
+  void set_spatial_scale(float value);
+
+  // @@protoc_insertion_point(class_scope:opencv_caffe.ROIPoolingParameter)
+ private:
+  inline void set_has_pooled_h();
+  inline void clear_has_pooled_h();
+  inline void set_has_pooled_w();
+  inline void clear_has_pooled_w();
+  inline void set_has_spatial_scale();
+  inline void clear_has_spatial_scale();
+
+  ::google::protobuf::internal::InternalMetadataWithArena _internal_metadata_;
+  ::google::protobuf::internal::HasBits<1> _has_bits_;
+  mutable int _cached_size_;
+  ::google::protobuf::uint32 pooled_h_;
+  ::google::protobuf::uint32 pooled_w_;
+  float spatial_scale_;
+  friend void  protobuf_InitDefaults_opencv_2dcaffe_2eproto_impl();
+  friend void  protobuf_AddDesc_opencv_2dcaffe_2eproto_impl();
+  friend void protobuf_AssignDesc_opencv_2dcaffe_2eproto();
+  friend void protobuf_ShutdownFile_opencv_2dcaffe_2eproto();
+
+  void InitAsDefaultInstance();
+};
+extern ::google::protobuf::internal::ExplicitlyConstructed<ROIPoolingParameter> ROIPoolingParameter_default_instance_;
+
 // ===================================================================
 
 
@@ -19015,15 +19146,60 @@ inline void LayerParameter::set_allocated_reshape_param(::opencv_caffe::ReshapeP
   // @@protoc_insertion_point(field_set_allocated:opencv_caffe.LayerParameter.reshape_param)
 }
 
+// optional .opencv_caffe.ROIPoolingParameter roi_pooling_param = 8266711;
+inline bool LayerParameter::has_roi_pooling_param() const {
+  return (_has_bits_[1] & 0x00200000u) != 0;
+}
+inline void LayerParameter::set_has_roi_pooling_param() {
+  _has_bits_[1] |= 0x00200000u;
+}
+inline void LayerParameter::clear_has_roi_pooling_param() {
+  _has_bits_[1] &= ~0x00200000u;
+}
+inline void LayerParameter::clear_roi_pooling_param() {
+  if (roi_pooling_param_ != NULL) roi_pooling_param_->::opencv_caffe::ROIPoolingParameter::Clear();
+  clear_has_roi_pooling_param();
+}
+inline const ::opencv_caffe::ROIPoolingParameter& LayerParameter::roi_pooling_param() const {
+  // @@protoc_insertion_point(field_get:opencv_caffe.LayerParameter.roi_pooling_param)
+  return roi_pooling_param_ != NULL ? *roi_pooling_param_
+                         : *::opencv_caffe::ROIPoolingParameter::internal_default_instance();
+}
+inline ::opencv_caffe::ROIPoolingParameter* LayerParameter::mutable_roi_pooling_param() {
+  set_has_roi_pooling_param();
+  if (roi_pooling_param_ == NULL) {
+    roi_pooling_param_ = new ::opencv_caffe::ROIPoolingParameter;
+  }
+  // @@protoc_insertion_point(field_mutable:opencv_caffe.LayerParameter.roi_pooling_param)
+  return roi_pooling_param_;
+}
+inline ::opencv_caffe::ROIPoolingParameter* LayerParameter::release_roi_pooling_param() {
+  // @@protoc_insertion_point(field_release:opencv_caffe.LayerParameter.roi_pooling_param)
+  clear_has_roi_pooling_param();
+  ::opencv_caffe::ROIPoolingParameter* temp = roi_pooling_param_;
+  roi_pooling_param_ = NULL;
+  return temp;
+}
+inline void LayerParameter::set_allocated_roi_pooling_param(::opencv_caffe::ROIPoolingParameter* roi_pooling_param) {
+  delete roi_pooling_param_;
+  roi_pooling_param_ = roi_pooling_param;
+  if (roi_pooling_param) {
+    set_has_roi_pooling_param();
+  } else {
+    clear_has_roi_pooling_param();
+  }
+  // @@protoc_insertion_point(field_set_allocated:opencv_caffe.LayerParameter.roi_pooling_param)
+}
+
 // optional .opencv_caffe.ScaleParameter scale_param = 142;
 inline bool LayerParameter::has_scale_param() const {
-  return (_has_bits_[1] & 0x00200000u) != 0;
+  return (_has_bits_[1] & 0x00400000u) != 0;
 }
 inline void LayerParameter::set_has_scale_param() {
-  _has_bits_[1] |= 0x00200000u;
+  _has_bits_[1] |= 0x00400000u;
 }
 inline void LayerParameter::clear_has_scale_param() {
-  _has_bits_[1] &= ~0x00200000u;
+  _has_bits_[1] &= ~0x00400000u;
 }
 inline void LayerParameter::clear_scale_param() {
   if (scale_param_ != NULL) scale_param_->::opencv_caffe::ScaleParameter::Clear();
@@ -19062,13 +19238,13 @@ inline void LayerParameter::set_allocated_scale_param(::opencv_caffe::ScaleParam
 
 // optional .opencv_caffe.SigmoidParameter sigmoid_param = 124;
 inline bool LayerParameter::has_sigmoid_param() const {
-  return (_has_bits_[1] & 0x00400000u) != 0;
+  return (_has_bits_[1] & 0x00800000u) != 0;
 }
 inline void LayerParameter::set_has_sigmoid_param() {
-  _has_bits_[1] |= 0x00400000u;
+  _has_bits_[1] |= 0x00800000u;
 }
 inline void LayerParameter::clear_has_sigmoid_param() {
-  _has_bits_[1] &= ~0x00400000u;
+  _has_bits_[1] &= ~0x00800000u;
 }
 inline void LayerParameter::clear_sigmoid_param() {
   if (sigmoid_param_ != NULL) sigmoid_param_->::opencv_caffe::SigmoidParameter::Clear();
@@ -19107,13 +19283,13 @@ inline void LayerParameter::set_allocated_sigmoid_param(::opencv_caffe::SigmoidP
 
 // optional .opencv_caffe.SoftmaxParameter softmax_param = 125;
 inline bool LayerParameter::has_softmax_param() const {
-  return (_has_bits_[1] & 0x00800000u) != 0;
+  return (_has_bits_[1] & 0x01000000u) != 0;
 }
 inline void LayerParameter::set_has_softmax_param() {
-  _has_bits_[1] |= 0x00800000u;
+  _has_bits_[1] |= 0x01000000u;
 }
 inline void LayerParameter::clear_has_softmax_param() {
-  _has_bits_[1] &= ~0x00800000u;
+  _has_bits_[1] &= ~0x01000000u;
 }
 inline void LayerParameter::clear_softmax_param() {
   if (softmax_param_ != NULL) softmax_param_->::opencv_caffe::SoftmaxParameter::Clear();
@@ -19152,13 +19328,13 @@ inline void LayerParameter::set_allocated_softmax_param(::opencv_caffe::SoftmaxP
 
 // optional .opencv_caffe.SPPParameter spp_param = 132;
 inline bool LayerParameter::has_spp_param() const {
-  return (_has_bits_[1] & 0x01000000u) != 0;
+  return (_has_bits_[1] & 0x02000000u) != 0;
 }
 inline void LayerParameter::set_has_spp_param() {
-  _has_bits_[1] |= 0x01000000u;
+  _has_bits_[1] |= 0x02000000u;
 }
 inline void LayerParameter::clear_has_spp_param() {
-  _has_bits_[1] &= ~0x01000000u;
+  _has_bits_[1] &= ~0x02000000u;
 }
 inline void LayerParameter::clear_spp_param() {
   if (spp_param_ != NULL) spp_param_->::opencv_caffe::SPPParameter::Clear();
@@ -19197,13 +19373,13 @@ inline void LayerParameter::set_allocated_spp_param(::opencv_caffe::SPPParameter
 
 // optional .opencv_caffe.SliceParameter slice_param = 126;
 inline bool LayerParameter::has_slice_param() const {
-  return (_has_bits_[1] & 0x02000000u) != 0;
+  return (_has_bits_[1] & 0x04000000u) != 0;
 }
 inline void LayerParameter::set_has_slice_param() {
-  _has_bits_[1] |= 0x02000000u;
+  _has_bits_[1] |= 0x04000000u;
 }
 inline void LayerParameter::clear_has_slice_param() {
-  _has_bits_[1] &= ~0x02000000u;
+  _has_bits_[1] &= ~0x04000000u;
 }
 inline void LayerParameter::clear_slice_param() {
   if (slice_param_ != NULL) slice_param_->::opencv_caffe::SliceParameter::Clear();
@@ -19242,13 +19418,13 @@ inline void LayerParameter::set_allocated_slice_param(::opencv_caffe::SliceParam
 
 // optional .opencv_caffe.TanHParameter tanh_param = 127;
 inline bool LayerParameter::has_tanh_param() const {
-  return (_has_bits_[1] & 0x04000000u) != 0;
+  return (_has_bits_[1] & 0x08000000u) != 0;
 }
 inline void LayerParameter::set_has_tanh_param() {
-  _has_bits_[1] |= 0x04000000u;
+  _has_bits_[1] |= 0x08000000u;
 }
 inline void LayerParameter::clear_has_tanh_param() {
-  _has_bits_[1] &= ~0x04000000u;
+  _has_bits_[1] &= ~0x08000000u;
 }
 inline void LayerParameter::clear_tanh_param() {
   if (tanh_param_ != NULL) tanh_param_->::opencv_caffe::TanHParameter::Clear();
@@ -19287,13 +19463,13 @@ inline void LayerParameter::set_allocated_tanh_param(::opencv_caffe::TanHParamet
 
 // optional .opencv_caffe.ThresholdParameter threshold_param = 128;
 inline bool LayerParameter::has_threshold_param() const {
-  return (_has_bits_[1] & 0x08000000u) != 0;
+  return (_has_bits_[1] & 0x10000000u) != 0;
 }
 inline void LayerParameter::set_has_threshold_param() {
-  _has_bits_[1] |= 0x08000000u;
+  _has_bits_[1] |= 0x10000000u;
 }
 inline void LayerParameter::clear_has_threshold_param() {
-  _has_bits_[1] &= ~0x08000000u;
+  _has_bits_[1] &= ~0x10000000u;
 }
 inline void LayerParameter::clear_threshold_param() {
   if (threshold_param_ != NULL) threshold_param_->::opencv_caffe::ThresholdParameter::Clear();
@@ -19332,13 +19508,13 @@ inline void LayerParameter::set_allocated_threshold_param(::opencv_caffe::Thresh
 
 // optional .opencv_caffe.TileParameter tile_param = 138;
 inline bool LayerParameter::has_tile_param() const {
-  return (_has_bits_[1] & 0x10000000u) != 0;
+  return (_has_bits_[1] & 0x20000000u) != 0;
 }
 inline void LayerParameter::set_has_tile_param() {
-  _has_bits_[1] |= 0x10000000u;
+  _has_bits_[1] |= 0x20000000u;
 }
 inline void LayerParameter::clear_has_tile_param() {
-  _has_bits_[1] &= ~0x10000000u;
+  _has_bits_[1] &= ~0x20000000u;
 }
 inline void LayerParameter::clear_tile_param() {
   if (tile_param_ != NULL) tile_param_->::opencv_caffe::TileParameter::Clear();
@@ -19377,13 +19553,13 @@ inline void LayerParameter::set_allocated_tile_param(::opencv_caffe::TileParamet
 
 // optional .opencv_caffe.WindowDataParameter window_data_param = 129;
 inline bool LayerParameter::has_window_data_param() const {
-  return (_has_bits_[1] & 0x20000000u) != 0;
+  return (_has_bits_[1] & 0x40000000u) != 0;
 }
 inline void LayerParameter::set_has_window_data_param() {
-  _has_bits_[1] |= 0x20000000u;
+  _has_bits_[1] |= 0x40000000u;
 }
 inline void LayerParameter::clear_has_window_data_param() {
-  _has_bits_[1] &= ~0x20000000u;
+  _has_bits_[1] &= ~0x40000000u;
 }
 inline void LayerParameter::clear_window_data_param() {
   if (window_data_param_ != NULL) window_data_param_->::opencv_caffe::WindowDataParameter::Clear();
@@ -28660,6 +28836,85 @@ inline void NormalizedBBox::set_size(float value) {
 inline const NormalizedBBox* NormalizedBBox::internal_default_instance() {
   return &NormalizedBBox_default_instance_.get();
 }
+// -------------------------------------------------------------------
+
+// ROIPoolingParameter
+
+// optional uint32 pooled_h = 1 [default = 0];
+inline bool ROIPoolingParameter::has_pooled_h() const {
+  return (_has_bits_[0] & 0x00000001u) != 0;
+}
+inline void ROIPoolingParameter::set_has_pooled_h() {
+  _has_bits_[0] |= 0x00000001u;
+}
+inline void ROIPoolingParameter::clear_has_pooled_h() {
+  _has_bits_[0] &= ~0x00000001u;
+}
+inline void ROIPoolingParameter::clear_pooled_h() {
+  pooled_h_ = 0u;
+  clear_has_pooled_h();
+}
+inline ::google::protobuf::uint32 ROIPoolingParameter::pooled_h() const {
+  // @@protoc_insertion_point(field_get:opencv_caffe.ROIPoolingParameter.pooled_h)
+  return pooled_h_;
+}
+inline void ROIPoolingParameter::set_pooled_h(::google::protobuf::uint32 value) {
+  set_has_pooled_h();
+  pooled_h_ = value;
+  // @@protoc_insertion_point(field_set:opencv_caffe.ROIPoolingParameter.pooled_h)
+}
+
+// optional uint32 pooled_w = 2 [default = 0];
+inline bool ROIPoolingParameter::has_pooled_w() const {
+  return (_has_bits_[0] & 0x00000002u) != 0;
+}
+inline void ROIPoolingParameter::set_has_pooled_w() {
+  _has_bits_[0] |= 0x00000002u;
+}
+inline void ROIPoolingParameter::clear_has_pooled_w() {
+  _has_bits_[0] &= ~0x00000002u;
+}
+inline void ROIPoolingParameter::clear_pooled_w() {
+  pooled_w_ = 0u;
+  clear_has_pooled_w();
+}
+inline ::google::protobuf::uint32 ROIPoolingParameter::pooled_w() const {
+  // @@protoc_insertion_point(field_get:opencv_caffe.ROIPoolingParameter.pooled_w)
+  return pooled_w_;
+}
+inline void ROIPoolingParameter::set_pooled_w(::google::protobuf::uint32 value) {
+  set_has_pooled_w();
+  pooled_w_ = value;
+  // @@protoc_insertion_point(field_set:opencv_caffe.ROIPoolingParameter.pooled_w)
+}
+
+// optional float spatial_scale = 3 [default = 1];
+inline bool ROIPoolingParameter::has_spatial_scale() const {
+  return (_has_bits_[0] & 0x00000004u) != 0;
+}
+inline void ROIPoolingParameter::set_has_spatial_scale() {
+  _has_bits_[0] |= 0x00000004u;
+}
+inline void ROIPoolingParameter::clear_has_spatial_scale() {
+  _has_bits_[0] &= ~0x00000004u;
+}
+inline void ROIPoolingParameter::clear_spatial_scale() {
+  spatial_scale_ = 1;
+  clear_has_spatial_scale();
+}
+inline float ROIPoolingParameter::spatial_scale() const {
+  // @@protoc_insertion_point(field_get:opencv_caffe.ROIPoolingParameter.spatial_scale)
+  return spatial_scale_;
+}
+inline void ROIPoolingParameter::set_spatial_scale(float value) {
+  set_has_spatial_scale();
+  spatial_scale_ = value;
+  // @@protoc_insertion_point(field_set:opencv_caffe.ROIPoolingParameter.spatial_scale)
+}
+
+inline const ROIPoolingParameter* ROIPoolingParameter::internal_default_instance() {
+  return &ROIPoolingParameter_default_instance_.get();
+}
 #endif  // !PROTOBUF_INLINE_NOT_IN_HEADERS
 // -------------------------------------------------------------------
 
@@ -28795,6 +29050,8 @@ inline const NormalizedBBox* NormalizedBBox::internal_default_instance() {
 
 // -------------------------------------------------------------------
 
+// -------------------------------------------------------------------
+
 
 // @@protoc_insertion_point(namespace_scope)
 

modules/dnn/src/caffe/opencv-caffe.proto

@@ -552,6 +552,7 @@ message LayerParameter {
   optional ReductionParameter reduction_param = 136;
   optional ReLUParameter relu_param = 123;
   optional ReshapeParameter reshape_param = 133;
+  optional ROIPoolingParameter roi_pooling_param = 8266711;  // https://github.com/rbgirshick/caffe-fast-rcnn/tree/fast-rcnn
   optional ScaleParameter scale_param = 142;
   optional SigmoidParameter sigmoid_param = 124;
   optional SoftmaxParameter softmax_param = 125;
@@ -1605,3 +1606,15 @@ message NormalizedBBox {
   optional float score = 7;
   optional float size = 8;
 }
+
+// origin: https://github.com/rbgirshick/caffe-fast-rcnn/tree/fast-rcnn
+// Message that stores parameters used by ROIPoolingLayer
+message ROIPoolingParameter {
+  // Pad, kernel size, and stride are all given as a single value for equal
+  // dimensions in height and width or as Y, X pairs.
+  optional uint32 pooled_h = 1 [default = 0]; // The pooled output height
+  optional uint32 pooled_w = 2 [default = 0]; // The pooled output width
+  // Multiplicative spatial scale factor to translate ROI coords from their
+  // input scale to the scale used when pooling
+  optional float spatial_scale = 3 [default = 1];
+}

modules/dnn/src/init.cpp

@@ -88,6 +88,7 @@ void initializeLayerFactory()
     CV_DNN_REGISTER_LAYER_CLASS(Convolution,    ConvolutionLayer);
     CV_DNN_REGISTER_LAYER_CLASS(Deconvolution,  DeconvolutionLayer);
     CV_DNN_REGISTER_LAYER_CLASS(Pooling,        PoolingLayer);
+    CV_DNN_REGISTER_LAYER_CLASS(ROIPooling,     PoolingLayer);
     CV_DNN_REGISTER_LAYER_CLASS(LRN,            LRNLayer);
     CV_DNN_REGISTER_LAYER_CLASS(InnerProduct,   InnerProductLayer);
     CV_DNN_REGISTER_LAYER_CLASS(Softmax,        SoftmaxLayer);

modules/dnn/src/layers/pooling_layer.cpp

@@ -65,6 +65,7 @@ public:
     {
         type = PoolingLayer::MAX;
         computeMaxIdx = true;
+        globalPooling = false;
 
         if (params.has("pool"))
         {
@@ -77,12 +78,18 @@ public:
                 type = PoolingLayer::STOCHASTIC;
             else
                 CV_Error(Error::StsBadArg, "Unknown pooling type \"" + pool + "\"");
+            getPoolingKernelParams(params, kernel.height, kernel.width, globalPooling,
+                                   pad.height, pad.width, stride.height, stride.width, padMode);
+        }
+        else if (params.has("pooled_w") || params.has("pooled_h") || params.has("spatial_scale"))
+        {
+            type = PoolingLayer::ROI;
         }
-
-        getPoolingKernelParams(params, kernel.height, kernel.width, globalPooling,
-                               pad.height, pad.width, stride.height, stride.width, padMode);
         setParamsFrom(params);
         ceilMode = params.get<bool>("ceil_mode", true);
+        pooledSize.width = params.get<uint32_t>("pooled_w", 1);
+        pooledSize.height = params.get<uint32_t>("pooled_h", 1);
+        spatialScale = params.get<float>("spatial_scale", 1);
     }
 
 #ifdef HAVE_OPENCL
@@ -91,7 +98,7 @@ public:
 
     void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs)
     {
-        CV_Assert(inputs.size() == 1);
+        CV_Assert(!inputs.empty());
 
         cv::Size inp(inputs[0]->size[3], inputs[0]->size[2]),
                 out(outputs[0].size[3], outputs[0].size[2]);
@@ -171,20 +178,23 @@ public:
         CV_TRACE_FUNCTION();
         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
-        for (size_t ii = 0; ii < inputs.size(); ii++)
+        switch (type)
         {
-            switch (type)
-            {
-                case MAX:
-                    maxPooling(*inputs[ii], outputs[2 * ii], outputs[2 * ii + 1]);
-                    break;
-                case AVE:
-                    avePooling(*inputs[ii], outputs[ii]);
-                    break;
-                default:
-                    CV_Error(Error::StsNotImplemented, "Not implemented");
-                    break;
-            }
+            case MAX:
+                CV_Assert(inputs.size() == 1, outputs.size() == 2);
+                maxPooling(*inputs[0], outputs[0], outputs[1]);
+                break;
+            case AVE:
+                CV_Assert(inputs.size() == 1, outputs.size() == 1);
+                avePooling(*inputs[0], outputs[0]);
+                break;
+            case ROI:
+                CV_Assert(inputs.size() == 2, outputs.size() == 1);
+                roiPooling(*inputs[0], *inputs[1], outputs[0]);
+                break;
+            default:
+                CV_Error(Error::StsNotImplemented, "Not implemented");
+                break;
         }
     }
 
@@ -201,29 +211,33 @@ public:
     class PoolingInvoker : public ParallelLoopBody
     {
     public:
-        const Mat* src;
+        const Mat* src, *rois;
         Mat *dst, *mask;
         Size kernel, stride, pad;
         int nstripes;
         bool computeMaxIdx;
         std::vector<int> ofsbuf;
         int poolingType;
+        float spatialScale;
 
-        PoolingInvoker() : src(0), dst(0), mask(0), nstripes(0), computeMaxIdx(0), poolingType(PoolingLayer::MAX) {}
+        PoolingInvoker() : src(0), rois(0), dst(0), mask(0), nstripes(0),
+                           computeMaxIdx(0), poolingType(PoolingLayer::MAX), spatialScale(0) {}
 
-        static void run(const Mat& src, Mat& dst, Mat& mask, Size kernel,
-                        Size stride, Size pad, int poolingType,
+        static void run(const Mat& src, const Mat& rois, Mat& dst, Mat& mask, Size kernel,
+                        Size stride, Size pad, int poolingType, float spatialScale,
                         bool computeMaxIdx, int nstripes)
         {
             CV_Assert(src.isContinuous() && dst.isContinuous() &&
                       src.type() == CV_32F && src.type() == dst.type() &&
                       src.dims == 4 && dst.dims == 4 &&
-                      src.size[0] == dst.size[0] && src.size[1] == dst.size[1] &&
+                      (poolingType == ROI && dst.size[0] == rois.size[0] ||
+                       src.size[0] == dst.size[0]) && src.size[1] == dst.size[1] &&
                       (mask.empty() || (mask.type() == src.type() && mask.size == dst.size)));
 
             PoolingInvoker p;
 
             p.src = &src;
+            p.rois = &rois;
             p.dst = &dst;
             p.mask = &mask;
             p.kernel = kernel;
@@ -232,6 +246,7 @@ public:
             p.nstripes = nstripes;
             p.computeMaxIdx = computeMaxIdx;
             p.poolingType = poolingType;
+            p.spatialScale = spatialScale;
 
             if( !computeMaxIdx )
             {
@@ -273,12 +288,39 @@ public:
                 ofs /= height;
                 int c = (int)(ofs % channels);
                 int n = (int)(ofs / channels);
-                int ystart = y0 * stride_h - pad_h;
-                int yend = min(ystart + kernel_h, inp_height + pad_h);
+                int ystart, yend;
+
+                const float *srcData;
+                int xstartROI = 0;
+                float roiRatio = 0;
+                if (poolingType == ROI)
+                {
+                    const float *roisData = rois->ptr<float>(n);
+                    int ystartROI = round(roisData[2] * spatialScale);
+                    int yendROI = round(roisData[4] * spatialScale);
+                    int roiHeight = std::max(yendROI - ystartROI + 1, 1);
+                    roiRatio = (float)roiHeight / height;
+
+                    ystart = ystartROI + y0 * roiRatio;
+                    yend = ystartROI + std::ceil((y0 + 1) * roiRatio);
+
+                    xstartROI = round(roisData[1] * spatialScale);
+                    int xendROI = round(roisData[3] * spatialScale);
+                    int roiWidth = std::max(xendROI - xstartROI + 1, 1);
+                    roiRatio = (float)roiWidth / width;
+
+                    CV_Assert(roisData[0] < src->size[0]);
+                    srcData = src->ptr<float>(roisData[0], c);
+                }
+                else
+                {
+                    ystart = y0 * stride_h - pad_h;
+                    yend = min(ystart + kernel_h, inp_height + pad_h);
+                    srcData = src->ptr<float>(n, c);
+                }
                 int ydelta = yend - ystart;
                 ystart = max(ystart, 0);
                 yend = min(yend, inp_height);
-                const float *srcData = src->ptr<float>(n, c);
                 float *dstData = dst->ptr<float>(n, c, y0);
                 float *dstMaskData = mask->data ? mask->ptr<float>(n, c, y0) : 0;
 
@@ -286,13 +328,29 @@ public:
                 ofs0 += delta;
                 int x1 = x0 + delta;
 
-                if( poolingType == PoolingLayer::MAX )
+                if( poolingType == MAX || poolingType == ROI)
                     for( ; x0 < x1; x0++ )
                     {
-                        int xstart = x0 * stride_w - pad_w;
-                        int xend = min(xstart + kernel_w, inp_width);
+                        int xstart, xend;
+                        if (poolingType == ROI)
+                        {
+                            xstart = xstartROI + x0 * roiRatio;
+                            xend = xstartROI + std::ceil((x0 + 1) * roiRatio);
+                        }
+                        else
+                        {
+                            xstart = x0 * stride_w - pad_w;
+                            xend = xstart + kernel_w;
+                        }
                         xstart = max(xstart, 0);
-
+                        xend = min(xend, inp_width);
+                        if (xstart >= xend || ystart >= yend)
+                        {
+                            dstData[x0] = 0;
+                            if (compMaxIdx && dstMaskData)
+                                dstMaskData[x0] = -1;
+                            continue;
+                        }
 #if CV_SIMD128
                         if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_w + kernel_w < inp_width )
                         {
@@ -489,14 +547,22 @@ public:
     void maxPooling(Mat &src, Mat &dst, Mat &mask)
     {
         const int nstripes = getNumThreads();
-        PoolingInvoker::run(src, dst, mask, kernel, stride, pad, type, computeMaxIdx, nstripes);
+        Mat rois;
+        PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
     }
 
     void avePooling(Mat &src, Mat &dst)
+    {
+        const int nstripes = getNumThreads();
+        Mat rois, mask;
+        PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
+    }
+
+    void roiPooling(const Mat &src, const Mat &rois, Mat &dst)
     {
         const int nstripes = getNumThreads();
         Mat mask;
-        PoolingInvoker::run(src, dst, mask, kernel, stride, pad, type, computeMaxIdx, nstripes);
+        PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
     }
 
     virtual Ptr<BackendNode> initMaxPoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
@@ -632,6 +698,11 @@ public:
             out.height = 1;
             out.width = 1;
         }
+        else if (type == PoolingLayer::ROI)
+        {
+            out.height = pooledSize.height;
+            out.width = pooledSize.width;
+        }
         else if (padMode.empty())
         {
             float height = (float)(in.height + 2 * pad.height - kernel.height) / stride.height;
@@ -656,17 +727,13 @@ public:
             getConvPoolOutParams(in, kernel, stride, padMode, Size(1, 1), out);
         }
 
-        outputs.resize(type == MAX ? 2 * inputs.size() : inputs.size());
-        for (size_t i = 0; i < inputs.size(); i++)
+        int dims[] = {inputs[0][0], inputs[0][1], out.height, out.width};
+        if (type == ROI)
         {
-            size_t index = type == MAX ? 2*i : i;
-            int dims[] = {inputs[i][0], inputs[i][1], out.height, out.width};
-            outputs[index] = shape(dims);
-
-            if (type == MAX)
-                outputs[index + 1] = shape(dims);
+            CV_Assert(inputs.size() == 2);
+            dims[0] = inputs[1][0];  // Number of proposals;
         }
-
+        outputs.assign(type == MAX ? 2 : 1, shape(dims));
         return false;
     }
 

modules/dnn/test/test_layers.cpp

@@ -560,4 +560,20 @@ TEST(Layer_Test_Reorg, Accuracy)
     testLayerUsingDarknetModels("reorg", false, false);
 }
 
+TEST(Layer_Test_ROIPooling, Accuracy)
+{
+    Net net = readNetFromCaffe(_tf("net_roi_pooling.prototxt"));
+
+    Mat inp = blobFromNPY(_tf("net_roi_pooling.input.npy"));
+    Mat rois = blobFromNPY(_tf("net_roi_pooling.rois.npy"));
+    Mat ref = blobFromNPY(_tf("net_roi_pooling.npy"));
+
+    net.setInput(inp, "input");
+    net.setInput(rois, "rois");
+
+    Mat out = net.forward();
+
+    normAssert(out, ref);
+}
+
 }