COMPMID-1999: Add support for GenerateProposals operator in CL

Change-Id: Ie08a6874347085f96b00f25bdb605eee7d683c25
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/719
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

13 files changed, 695 insertions, 17 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index a7d371dabc..4ecb885440 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -307,6 +307,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemmlowp_output_stage_quantize_down", "gemmlowp.cl" },
     { "gemmlowp_output_stage_quantize_down_fixedpoint", "gemmlowp.cl" },
     { "gemmlowp_output_stage_quantize_down_float", "gemmlowp.cl" },
+    { "generate_proposals_compute_all_anchors", "generate_proposals.cl" },
     { "harris_score_3x3", "harris_corners.cl" },
     { "harris_score_5x5", "harris_corners.cl" },
     { "harris_score_7x7", "harris_corners.cl" },
@@ -706,6 +707,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/gemv.clembed"
     },
     {
+        "generate_proposals.cl",
+#include "./cl_kernels/generate_proposals.clembed"
+    },
+    {
         "harris_corners.cl",
 #include "./cl_kernels/harris_corners.clembed"
     },
diff --git a/src/core/CL/cl_kernels/bounding_box_transform.cl b/src/core/CL/cl_kernels/bounding_box_transform.cl
index 77db5d9311..e6f470a962 100644
--- a/src/core/CL/cl_kernels/bounding_box_transform.cl
+++ b/src/core/CL/cl_kernels/bounding_box_transform.cl
@@ -28,11 +28,11 @@
 /** Perform a padded copy of input tensor to the output tensor. Padding values are defined at compile time
  *
  * @attention The following variables must be passed at compile time:
- * -# -DDATA_TYPE = Tensor data type. Supported data types: F16/F32
+ * -# -DDATA_TYPE= Tensor data type. Supported data types: F16/F32
  * -# -DWEIGHT{X,Y,W,H}= Weights [wx, wy, ww, wh] for the deltas
  * -# -DIMG_WIDTH= Original image width
  * -# -DIMG_HEIGHT= Original image height
- * -# -DBOX_FIELDS=Number of fields that are used to represent a box in boxes
+ * -# -DBOX_FIELDS= Number of fields that are used to represent a box in boxes
  *
  * @param[in]  boxes_ptr                                Pointer to the boxes tensor. Supported data types: F16/F32
  * @param[in]  boxes_stride_x                           Stride of the boxes tensor in X dimension (in bytes)
diff --git a/src/core/CL/cl_kernels/generate_proposals.cl b/src/core/CL/cl_kernels/generate_proposals.cl
new file mode 100644
index 0000000000..a947dad523
--- /dev/null
+++ b/src/core/CL/cl_kernels/generate_proposals.cl
@@ -0,0 +1,88 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "helpers.h"
+
+/** Generate all the region of interests based on the image size and the anchors passed in. For each element (x,y) of the
+ * grid, it will generate NUM_ANCHORS rois, given by shifting the grid position to match the anchor.
+ *
+ * @attention The following variables must be passed at compile time:
+ * -# -DDATA_TYPE= Tensor data type. Supported data types: F16/F32
+ * -# -DHEIGHT= Height of the feature map on which this kernel is applied
+ * -# -DWIDTH= Width of the feature map on which this kernel is applied
+ * -# -DNUM_ANCHORS= Number of anchors to be used to generate the rois per each pixel
+ * -# -DSTRIDE= Stride to be applied at each different pixel position (i.e., x_range = (1:WIDTH)*STRIDE and y_range = (1:HEIGHT)*STRIDE
+ * -# -DNUM_ROI_FIELDS= Number of fields used to represent a roi
+ *
+ * @param[in]  anchors_ptr                           Pointer to the anchors tensor. Supported data types: F16/F32
+ * @param[in]  anchors_stride_x                      Stride of the anchors tensor in X dimension (in bytes)
+ * @param[in]  anchors_step_x                        anchors_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  anchors_stride_y                      Stride of the anchors tensor in Y dimension (in bytes)
+ * @param[in]  anchors_step_y                        anchors_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  anchors_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  anchors_step_z                        anchors_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  anchors_offset_first_element_in_bytes The offset of the first element in the boxes tensor
+ * @param[out] rois_ptr                              Pointer to the rois. Supported data types: same as @p in_ptr
+ * @param[out] rois_stride_x                         Stride of the rois in X dimension (in bytes)
+ * @param[out] rois_step_x                           pred_boxes_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[out] rois_stride_y                         Stride of the rois in Y dimension (in bytes)
+ * @param[out] rois_step_y                           pred_boxes_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[out] rois_stride_z                         Stride of the rois in Z dimension (in bytes)
+ * @param[out] rois_step_z                           pred_boxes_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[out] rois_offset_first_element_in_bytes    The offset of the first element in the rois
+ */
+#if defined(DATA_TYPE) && defined(WIDTH) && defined(HEIGHT) && defined(NUM_ANCHORS) && defined(STRIDE) && defined(NUM_ROI_FIELDS)
+__kernel void generate_proposals_compute_all_anchors(
+    VECTOR_DECLARATION(anchors),
+    VECTOR_DECLARATION(rois))
+{
+    Vector anchors = CONVERT_TO_VECTOR_STRUCT_NO_STEP(anchors);
+    Vector rois    = CONVERT_TO_VECTOR_STRUCT(rois);
+
+    const size_t idx = get_global_id(0);
+    // Find the index of the anchor
+    const size_t anchor_idx = idx % NUM_ANCHORS;
+
+    // Find which shift is this thread using
+    const size_t shift_idx = idx / NUM_ANCHORS;
+
+    // Compute the shift on the X and Y direction (the shift depends exclusively by the index thread id)
+    const DATA_TYPE
+    shift_x = (DATA_TYPE)(shift_idx % WIDTH) * STRIDE;
+    const DATA_TYPE
+    shift_y = (DATA_TYPE)(shift_idx / WIDTH) * STRIDE;
+
+    const VEC_DATA_TYPE(DATA_TYPE, NUM_ROI_FIELDS)
+    shift = (VEC_DATA_TYPE(DATA_TYPE, NUM_ROI_FIELDS))(shift_x, shift_y, shift_x, shift_y);
+
+    // Read the given anchor
+    const VEC_DATA_TYPE(DATA_TYPE, NUM_ROI_FIELDS)
+    anchor = vload4(0, (__global DATA_TYPE *)vector_offset(&anchors, anchor_idx * NUM_ROI_FIELDS));
+
+    // Apply the shift to the anchor
+    const VEC_DATA_TYPE(DATA_TYPE, NUM_ROI_FIELDS)
+    shifted_anchor = anchor + shift;
+
+    vstore4(shifted_anchor, 0, (__global DATA_TYPE *)rois.ptr);
+}
+#endif //defined(DATA_TYPE) && defined(WIDTH) && defined(HEIGHT) && defined(NUM_ANCHORS) && defined(STRIDE) && defined(NUM_ROI_FIELDS)
diff --git a/src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp b/src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp
new file mode 100644
index 0000000000..f16422f815
--- /dev/null
+++ b/src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp
@@ -0,0 +1,128 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLGenerateProposalsLayerKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLArray.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Window.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(anchors, all_anchors);
+    ARM_COMPUTE_RETURN_ERROR_ON(anchors->dimension(0) != info.values_per_roi());
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(anchors, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(anchors->num_dimensions() > 2);
+    if(all_anchors->total_size() > 0)
+    {
+        size_t feature_height = info.feat_height();
+        size_t feature_width  = info.feat_width();
+        size_t num_anchors    = anchors->dimension(1);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(all_anchors, anchors);
+        ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->num_dimensions() > 2);
+        ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(0) != info.values_per_roi());
+        ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(1) != feature_height * feature_width * num_anchors);
+    }
+    return Status{};
+}
+} // namespace
+
+CLComputeAllAnchorsKernel::CLComputeAllAnchorsKernel()
+    : _anchors(nullptr), _all_anchors(nullptr)
+{
+}
+
+void CLComputeAllAnchorsKernel::configure(const ICLTensor *anchors, ICLTensor *all_anchors, const ComputeAnchorsInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(anchors, all_anchors);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(anchors->info(), all_anchors->info(), info));
+
+    // Metadata
+    const size_t   num_anchors = anchors->info()->dimension(1);
+    const DataType data_type   = anchors->info()->data_type();
+    const float    width       = info.feat_width();
+    const float    height      = info.feat_height();
+
+    // Initialize the output if empty
+    const TensorShape output_shape(info.values_per_roi(), width * height * num_anchors);
+    auto_init_if_empty(*all_anchors->info(), output_shape, 1, data_type);
+
+    // Set instance variables
+    _anchors     = anchors;
+    _all_anchors = all_anchors;
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
+    build_opts.add_option("-DWIDTH=" + float_to_string_with_full_precision(width));
+    build_opts.add_option("-DHEIGHT=" + float_to_string_with_full_precision(height));
+    build_opts.add_option("-DSTRIDE=" + float_to_string_with_full_precision(1.f / info.spatial_scale()));
+    build_opts.add_option("-DNUM_ANCHORS=" + support::cpp11::to_string(num_anchors));
+    build_opts.add_option("-DNUM_ROI_FIELDS=" + support::cpp11::to_string(info.values_per_roi()));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("generate_proposals_compute_all_anchors", build_opts.options()));
+
+    // The tensor all_anchors can be interpreted as an array of structs (each structs has values_per_roi fields).
+    // This means we don't need to pad on the X dimension, as we know in advance how many fields
+    // compose the struct.
+    Window win = calculate_max_window(*all_anchors->info(), Steps(info.values_per_roi()));
+    ICLKernel::configure_internal(win);
+}
+
+Status CLComputeAllAnchorsKernel::validate(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(anchors, all_anchors, info));
+    return Status{};
+}
+
+void CLComputeAllAnchorsKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    // Collapse everything on the first dimension
+    Window collapsed = window.collapse(ICLKernel::window(), Window::DimX);
+
+    // Set arguments
+    unsigned int idx = 0;
+    add_1D_tensor_argument(idx, _anchors, collapsed);
+    add_1D_tensor_argument(idx, _all_anchors, collapsed);
+
+    // Note that we don't need to loop over the slices, as we are launching exactly
+    // as many threads as all the anchors generated
+    enqueue(queue, *this, collapsed);
+}
+} // namespace arm_compute
diff --git a/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp b/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp
index 5e4b80aa5a..02150ff275 100644
--- a/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp
+++ b/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp
@@ -54,7 +54,7 @@ std::vector<int> SoftNMS(const ITensor *proposals, std::vector<std::vector<T>> &
         areas[i] = (x2[i] - x1[i] + 1.0) * (y2[i] - y1[i] + 1.0);
     }
 
-    // Note: Soft NMS scores have already been initialize with input scores
+    // Note: Soft NMS scores have already been initialized with input scores
 
     while(!inds.empty())
     {
@@ -150,17 +150,21 @@ std::vector<int> NonMaximaSuppression(const ITensor *proposals, std::vector<int>
 
         for(unsigned int j = 0; j < sorted_indices_temp.size(); ++j)
         {
-            const auto xx1 = std::max(x1[sorted_indices_temp.at(j)], x1[i]);
-            const auto yy1 = std::max(y1[sorted_indices_temp.at(j)], y1[i]);
-            const auto xx2 = std::min(x2[sorted_indices_temp.at(j)], x2[i]);
-            const auto yy2 = std::min(y2[sorted_indices_temp.at(j)], y2[i]);
-
-            const auto w     = std::max((xx2 - xx1 + 1.f), 0.f);
-            const auto h     = std::max((yy2 - yy1 + 1.f), 0.f);
-            const auto inter = w * h;
-            const auto ovr   = inter / (areas[i] + areas[sorted_indices_temp.at(j)] - inter);
-
-            if(ovr <= info.nms())
+            const float xx1 = std::max(x1[sorted_indices_temp.at(j)], x1[i]);
+            const float yy1 = std::max(y1[sorted_indices_temp.at(j)], y1[i]);
+            const float xx2 = std::min(x2[sorted_indices_temp.at(j)], x2[i]);
+            const float yy2 = std::min(y2[sorted_indices_temp.at(j)], y2[i]);
+
+            const float w     = std::max((xx2 - xx1 + 1.f), 0.f);
+            const float h     = std::max((yy2 - yy1 + 1.f), 0.f);
+            const float inter = w * h;
+            const float ovr   = inter / (areas[i] + areas[sorted_indices_temp.at(j)] - inter);
+            const float ctr_x = xx1 + (w / 2);
+            const float ctr_y = yy1 + (h / 2);
+
+            // If suppress_size is specified, filter the boxes based on their size and position
+            const bool keep_size = !info.suppress_size() || (w >= info.min_size() && h >= info.min_size() && ctr_x < info.im_width() && ctr_y < info.im_height());
+            if(ovr <= info.nms() && keep_size)
             {
                 new_indices.push_back(j);
             }
@@ -214,8 +218,9 @@ void CPPBoxWithNonMaximaSuppressionLimitKernel::run_nmslimit()
     for(int b = 0; b < batch_size; ++b)
     {
         const int num_boxes = _batch_splits_in == nullptr ? 1 : static_cast<int>(*reinterpret_cast<T *>(_batch_splits_in->ptr_to_element(Coordinates(b))));
-        // Skip first class
-        for(int j = 1; j < num_classes; ++j)
+        // Skip first class if there is more than 1 except if the number of classes is 1.
+        const int j_start = (num_classes == 1 ? 0 : 1);
+        for(int j = j_start; j < num_classes; ++j)
         {
             std::vector<T>   cur_scores(scores_count);
             std::vector<int> inds;
@@ -290,7 +295,7 @@ void CPPBoxWithNonMaximaSuppressionLimitKernel::run_nmslimit()
 
         // Write results
         int cur_out_idx = 0;
-        for(int j = 1; j < num_classes; ++j)
+        for(int j = j_start; j < num_classes; ++j)
         {
             auto     &cur_keep        = keeps[j];
             auto      cur_out_scores  = reinterpret_cast<T *>(_scores_out->ptr_to_element(Coordinates(cur_start_idx + cur_out_idx)));
diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index cac1a37099..a944d2c25d 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -448,6 +448,22 @@ NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, Node
     return fc_nid;
 }
 
+NodeID GraphBuilder::add_generate_proposals_node(Graph &g, NodeParams params, NodeIdxPair scores, NodeIdxPair deltas, NodeIdxPair anchors, GenerateProposalsInfo info)
+{
+    CHECK_NODEIDX_PAIR(scores, g);
+    CHECK_NODEIDX_PAIR(deltas, g);
+    CHECK_NODEIDX_PAIR(anchors, g);
+
+    NodeID nid = g.add_node<GenerateProposalsLayerNode>(info);
+
+    g.add_connection(scores.node_id, scores.index, nid, 0);
+    g.add_connection(deltas.node_id, deltas.index, nid, 1);
+    g.add_connection(anchors.node_id, anchors.index, nid, 2);
+
+    set_node_params(g, nid, params);
+    return nid;
+}
+
 NodeID GraphBuilder::add_normalization_node(Graph &g, NodeParams params, NodeIdxPair input, NormalizationLayerInfo norm_info)
 {
     return create_simple_single_input_output_node<NormalizationLayerNode>(g, params, input, norm_info);
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index 88d8e3c6c5..b9e3ddc0a3 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -192,6 +192,8 @@ std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &
             return detail::create_flatten_layer<CLFlattenLayer, CLTargetInfo>(*polymorphic_downcast<FlattenLayerNode *>(node));
         case NodeType::FullyConnectedLayer:
             return detail::create_fully_connected_layer<CLFullyConnectedLayer, CLTargetInfo>(*polymorphic_downcast<FullyConnectedLayerNode *>(node), ctx);
+        case NodeType::GenerateProposalsLayer:
+            return detail::create_generate_proposals_layer<CLGenerateProposalsLayer, CLTargetInfo>(*polymorphic_downcast<GenerateProposalsLayerNode *>(node), ctx);
         case NodeType::NormalizationLayer:
             return detail::create_normalization_layer<CLNormalizationLayer, CLTargetInfo>(*polymorphic_downcast<NormalizationLayerNode *>(node), ctx);
         case NodeType::NormalizePlanarYUVLayer:
diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp
index ca327c9771..4b71837a49 100644
--- a/src/graph/backends/CL/CLNodeValidator.cpp
+++ b/src/graph/backends/CL/CLNodeValidator.cpp
@@ -62,6 +62,8 @@ Status CLNodeValidator::validate(INode *node)
                    CLDepthwiseConvolutionLayer3x3>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::validate_detection_output_layer<CPPDetectionOutputLayer>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
+        case NodeType::GenerateProposalsLayer:
+            return detail::validate_generate_proposals_layer<CLGenerateProposalsLayer>(*polymorphic_downcast<GenerateProposalsLayerNode *>(node));
         case NodeType::NormalizePlanarYUVLayer:
             return detail::validate_normalize_planar_yuv_layer<CLNormalizePlanarYUVLayer>(*polymorphic_downcast<NormalizePlanarYUVLayerNode *>(node));
         case NodeType::PadLayer:
diff --git a/src/graph/backends/GLES/GCNodeValidator.cpp b/src/graph/backends/GLES/GCNodeValidator.cpp
index aaa031dbb9..f15ede6e2c 100644
--- a/src/graph/backends/GLES/GCNodeValidator.cpp
+++ b/src/graph/backends/GLES/GCNodeValidator.cpp
@@ -115,6 +115,8 @@ Status GCNodeValidator::validate(INode *node)
             return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : DetectionOutputLayer");
         case NodeType::FlattenLayer:
             return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : FlattenLayer");
+        case NodeType::GenerateProposalsLayer:
+            return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : GenerateProposalsLayer");
         case NodeType::NormalizePlanarYUVLayer:
             return detail::validate_normalize_planar_yuv_layer<GCNormalizePlanarYUVLayer>(*polymorphic_downcast<NormalizePlanarYUVLayerNode *>(node));
         case NodeType::PadLayer:
diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp
index 96236b66c3..b0feec563b 100644
--- a/src/graph/backends/NEON/NENodeValidator.cpp
+++ b/src/graph/backends/NEON/NENodeValidator.cpp
@@ -62,6 +62,8 @@ Status NENodeValidator::validate(INode *node)
                    NEDepthwiseConvolutionLayer3x3>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::validate_detection_output_layer<CPPDetectionOutputLayer>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
+        case NodeType::GenerateProposalsLayer:
+            return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : GenerateProposalsLayer");
         case NodeType::NormalizePlanarYUVLayer:
             return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : NormalizePlanarYUVLayer");
         case NodeType::PadLayer:
diff --git a/src/graph/nodes/GenerateProposalsLayerNode.cpp b/src/graph/nodes/GenerateProposalsLayerNode.cpp
new file mode 100644
index 0000000000..dabfc5aa10
--- /dev/null
+++ b/src/graph/nodes/GenerateProposalsLayerNode.cpp
@@ -0,0 +1,102 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/graph/nodes/GenerateProposalsLayerNode.h"
+
+#include "arm_compute/graph/Graph.h"
+#include "arm_compute/graph/INodeVisitor.h"
+
+#include "arm_compute/core/Helpers.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+GenerateProposalsLayerNode::GenerateProposalsLayerNode(GenerateProposalsInfo &info)
+    : _info(info)
+{
+    _input_edges.resize(3, EmptyEdgeID);
+    _outputs.resize(3, NullTensorID);
+}
+
+const GenerateProposalsInfo &GenerateProposalsLayerNode::info() const
+{
+    return _info;
+}
+
+bool GenerateProposalsLayerNode::forward_descriptors()
+{
+    if((input_id(0) != NullTensorID) && (input_id(1) != NullTensorID) && (input_id(2) != NullTensorID) && (output_id(0) != NullTensorID) && (output_id(1) != NullTensorID)
+       && (output_id(2) != NullTensorID))
+    {
+        for(unsigned int i = 0; i < 3; ++i)
+        {
+            Tensor *dst = output(i);
+            ARM_COMPUTE_ERROR_ON(dst == nullptr);
+            dst->desc() = configure_output(i);
+        }
+        return true;
+    }
+    return false;
+}
+
+TensorDescriptor GenerateProposalsLayerNode::configure_output(size_t idx) const
+{
+    ARM_COMPUTE_ERROR_ON(idx > 3);
+
+    const Tensor *src = input(0);
+    ARM_COMPUTE_ERROR_ON(src == nullptr);
+    TensorDescriptor output_desc = src->desc();
+
+    switch(idx)
+    {
+        case 0:
+            // Configure proposals output
+            output_desc.shape = TensorShape(5, src->desc().shape.total_size());
+            break;
+        case 1:
+            // Configure scores_out output
+            output_desc.shape = TensorShape(src->desc().shape.total_size());
+            break;
+        case 2:
+            // Configure num_valid_proposals
+            output_desc.shape     = TensorShape(1);
+            output_desc.data_type = DataType::U32;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported output index");
+    }
+    return output_desc;
+}
+
+NodeType GenerateProposalsLayerNode::type() const
+{
+    return NodeType::GenerateProposalsLayer;
+}
+
+void GenerateProposalsLayerNode::accept(INodeVisitor &v)
+{
+    v.visit(*this);
+}
+} // namespace graph
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLComputeAllAnchors.cpp b/src/runtime/CL/functions/CLComputeAllAnchors.cpp
new file mode 100644
index 0000000000..24c152f4d6
--- /dev/null
+++ b/src/runtime/CL/functions/CLComputeAllAnchors.cpp
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/CL/functions/CLComputeAllAnchors.h"
+
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+void CLComputeAllAnchors::configure(const ICLTensor *anchors, ICLTensor *all_anchors, const ComputeAnchorsInfo &info)
+{
+    // Configure ComputeAllAnchors kernel
+    auto k = arm_compute::support::cpp14::make_unique<CLComputeAllAnchorsKernel>();
+    k->configure(anchors, all_anchors, info);
+    _kernel = std::move(k);
+}
+
+Status CLComputeAllAnchors::validate(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info)
+{
+    return CLComputeAllAnchorsKernel::validate(anchors, all_anchors, info);
+}
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp b/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp
new file mode 100644
index 0000000000..c50132ea04
--- /dev/null
+++ b/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp
@@ -0,0 +1,284 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/CL/functions/CLGenerateProposalsLayer.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Types.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+CLGenerateProposalsLayer::CLGenerateProposalsLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)),
+      _permute_deltas_kernel(),
+      _flatten_deltas_kernel(),
+      _permute_scores_kernel(),
+      _flatten_scores_kernel(),
+      _compute_anchors_kernel(),
+      _bounding_box_kernel(),
+      _memset_kernel(),
+      _padded_copy_kernel(),
+      _cpp_nms_kernel(),
+      _is_nhwc(false),
+      _deltas_permuted(),
+      _deltas_flattened(),
+      _scores_permuted(),
+      _scores_flattened(),
+      _all_anchors(),
+      _all_proposals(),
+      _keeps_nms_unused(),
+      _classes_nms_unused(),
+      _proposals_4_roi_values(),
+      _num_valid_proposals(nullptr),
+      _scores_out(nullptr)
+{
+}
+
+void CLGenerateProposalsLayer::configure(const ICLTensor *scores, const ICLTensor *deltas, const ICLTensor *anchors, ICLTensor *proposals, ICLTensor *scores_out, ICLTensor *num_valid_proposals,
+                                         const GenerateProposalsInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals);
+    ARM_COMPUTE_ERROR_THROW_ON(CLGenerateProposalsLayer::validate(scores->info(), deltas->info(), anchors->info(), proposals->info(), scores_out->info(), num_valid_proposals->info(), info));
+
+    _is_nhwc                         = scores->info()->data_layout() == DataLayout::NHWC;
+    const DataType data_type         = deltas->info()->data_type();
+    const int      num_anchors       = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::CHANNEL));
+    const int      feat_width        = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::WIDTH));
+    const int      feat_height       = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::HEIGHT));
+    const int      total_num_anchors = num_anchors * feat_width * feat_height;
+    const int      pre_nms_topN      = info.pre_nms_topN();
+    const int      post_nms_topN     = info.post_nms_topN();
+    const size_t   values_per_roi    = info.values_per_roi();
+
+    // Compute all the anchors
+    _memory_group.manage(&_all_anchors);
+    _compute_anchors_kernel.configure(anchors, &_all_anchors, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale()));
+
+    const TensorShape flatten_shape_deltas(values_per_roi, total_num_anchors);
+    _deltas_flattened.allocator()->init(TensorInfo(flatten_shape_deltas, 1, data_type));
+
+    // Permute and reshape deltas
+    if(!_is_nhwc)
+    {
+        _memory_group.manage(&_deltas_permuted);
+        _memory_group.manage(&_deltas_flattened);
+        _permute_deltas_kernel.configure(deltas, &_deltas_permuted, PermutationVector{ 2, 0, 1 });
+        _flatten_deltas_kernel.configure(&_deltas_permuted, &_deltas_flattened);
+        _deltas_permuted.allocator()->allocate();
+    }
+    else
+    {
+        _memory_group.manage(&_deltas_flattened);
+        _flatten_deltas_kernel.configure(deltas, &_deltas_flattened);
+    }
+
+    const TensorShape flatten_shape_scores(1, total_num_anchors);
+    _scores_flattened.allocator()->init(TensorInfo(flatten_shape_scores, 1, data_type));
+
+    // Permute and reshape scores
+    if(!_is_nhwc)
+    {
+        _memory_group.manage(&_scores_permuted);
+        _memory_group.manage(&_scores_flattened);
+        _permute_scores_kernel.configure(scores, &_scores_permuted, PermutationVector{ 2, 0, 1 });
+        _flatten_scores_kernel.configure(&_scores_permuted, &_scores_flattened);
+        _scores_permuted.allocator()->allocate();
+    }
+    else
+    {
+        _memory_group.manage(&_scores_flattened);
+        _flatten_scores_kernel.configure(scores, &_scores_flattened);
+    }
+
+    // Bounding box transform
+    _memory_group.manage(&_all_proposals);
+    BoundingBoxTransformInfo bbox_info(info.im_width(), info.im_height(), 1.f);
+    _bounding_box_kernel.configure(&_all_anchors, &_all_proposals, &_deltas_flattened, bbox_info);
+    _deltas_flattened.allocator()->allocate();
+    _all_anchors.allocator()->allocate();
+
+    // The original layer implementation first selects the best pre_nms_topN anchors (thus having a lightweight sort)
+    // that are then transformed by bbox_transform. The boxes generated are then fed into a non-sorting NMS operation.
+    // Since we are reusing the NMS layer and we don't implement any CL/sort, we let NMS do the sorting (of all the input)
+    // and the filtering
+    const int   scores_nms_size = std::min<int>(std::min<int>(post_nms_topN, pre_nms_topN), total_num_anchors);
+    const float min_size_scaled = info.min_size() * info.im_scale();
+    _memory_group.manage(&_classes_nms_unused);
+    _memory_group.manage(&_keeps_nms_unused);
+
+    // Note that NMS needs outputs preinitialized.
+    auto_init_if_empty(*scores_out->info(), TensorShape(scores_nms_size), 1, data_type);
+    auto_init_if_empty(*_proposals_4_roi_values.info(), TensorShape(values_per_roi, scores_nms_size), 1, data_type);
+    auto_init_if_empty(*num_valid_proposals->info(), TensorShape(1), 1, DataType::U32);
+
+    // Initialize temporaries (unused) outputs
+    _classes_nms_unused.allocator()->init(TensorInfo(TensorShape(1, 1), 1, data_type));
+    _keeps_nms_unused.allocator()->init(*scores_out->info());
+
+    // Save the output (to map and unmap them at run)
+    _scores_out          = scores_out;
+    _num_valid_proposals = num_valid_proposals;
+
+    _memory_group.manage(&_proposals_4_roi_values);
+    _cpp_nms_kernel.configure(&_scores_flattened, &_all_proposals, nullptr, scores_out, &_proposals_4_roi_values, &_classes_nms_unused, nullptr, &_keeps_nms_unused, num_valid_proposals,
+                              BoxNMSLimitInfo(0.0f, info.nms_thres(), scores_nms_size, false, NMSType::LINEAR, 0.5f, 0.001f, true, min_size_scaled, info.im_width(), info.im_height()));
+    _keeps_nms_unused.allocator()->allocate();
+    _classes_nms_unused.allocator()->allocate();
+    _all_proposals.allocator()->allocate();
+    _scores_flattened.allocator()->allocate();
+
+    // Add the first column that represents the batch id. This will be all zeros, as we don't support multiple images
+    _padded_copy_kernel.configure(&_proposals_4_roi_values, proposals, PaddingList{ { 1, 0 } });
+    _proposals_4_roi_values.allocator()->allocate();
+
+    _memset_kernel.configure(proposals, PixelValue());
+}
+
+Status CLGenerateProposalsLayer::validate(const ITensorInfo *scores, const ITensorInfo *deltas, const ITensorInfo *anchors, const ITensorInfo *proposals, const ITensorInfo *scores_out,
+                                          const ITensorInfo *num_valid_proposals, const GenerateProposalsInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(scores, DataLayout::NCHW, DataLayout::NHWC);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(scores, deltas);
+
+    const int num_anchors       = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::CHANNEL));
+    const int feat_width        = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::WIDTH));
+    const int feat_height       = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::HEIGHT));
+    const int num_images        = scores->dimension(3);
+    const int total_num_anchors = num_anchors * feat_width * feat_height;
+    const int values_per_roi    = info.values_per_roi();
+
+    ARM_COMPUTE_RETURN_ERROR_ON(num_images > 1);
+
+    TensorInfo all_anchors_info(anchors->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLComputeAllAnchorsKernel::validate(anchors, &all_anchors_info, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale())));
+
+    TensorInfo deltas_permuted_info = deltas->clone()->set_tensor_shape(TensorShape(values_per_roi * num_anchors, feat_width, feat_height)).set_is_resizable(true);
+    TensorInfo scores_permuted_info = scores->clone()->set_tensor_shape(TensorShape(num_anchors, feat_width, feat_height)).set_is_resizable(true);
+    if(scores->data_layout() == DataLayout::NHWC)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(deltas, &deltas_permuted_info);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(scores, &scores_permuted_info);
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPermuteKernel::validate(deltas, &deltas_permuted_info, PermutationVector{ 2, 0, 1 }));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPermuteKernel::validate(scores, &scores_permuted_info, PermutationVector{ 2, 0, 1 }));
+    }
+
+    TensorInfo deltas_flattened_info(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(&deltas_permuted_info, &deltas_flattened_info));
+
+    TensorInfo scores_flattened_info(deltas->clone()->set_tensor_shape(TensorShape(1, total_num_anchors)).set_is_resizable(true));
+    TensorInfo proposals_4_roi_values(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
+
+    ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(&scores_permuted_info, &scores_flattened_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::validate(&all_anchors_info, &proposals_4_roi_values, &deltas_flattened_info, BoundingBoxTransformInfo(info.im_width(), info.im_height(),
+                                                                       1.f)));
+
+    ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(&proposals_4_roi_values, proposals, PaddingList{ { 0, 1 } }));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLMemsetKernel::validate(proposals, PixelValue()));
+
+    if(num_valid_proposals->total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(num_valid_proposals->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(num_valid_proposals->dimension(0) > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(num_valid_proposals, 1, DataType::U32);
+    }
+
+    if(proposals->total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(proposals->num_dimensions() > 2);
+        ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(0) != size_t(values_per_roi) + 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(1) != size_t(total_num_anchors));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(proposals, deltas);
+    }
+
+    if(scores_out->total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(scores_out->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(scores_out->dimension(0) != size_t(total_num_anchors));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(scores_out, scores);
+    }
+
+    return Status{};
+}
+
+void CLGenerateProposalsLayer::run_cpp_nms_kernel()
+{
+    // Map inputs
+    _scores_flattened.map(true);
+    _all_proposals.map(true);
+
+    // Map outputs
+    _scores_out->map(CLScheduler::get().queue(), true);
+    _proposals_4_roi_values.map(CLScheduler::get().queue(), true);
+    _num_valid_proposals->map(CLScheduler::get().queue(), true);
+    _keeps_nms_unused.map(true);
+    _classes_nms_unused.map(true);
+
+    // Run nms
+    CPPScheduler::get().schedule(&_cpp_nms_kernel, Window::DimX);
+
+    // Unmap outputs
+    _keeps_nms_unused.unmap();
+    _classes_nms_unused.unmap();
+    _scores_out->unmap(CLScheduler::get().queue());
+    _proposals_4_roi_values.unmap(CLScheduler::get().queue());
+    _num_valid_proposals->unmap(CLScheduler::get().queue());
+
+    // Unmap inputs
+    _scores_flattened.unmap();
+    _all_proposals.unmap();
+}
+
+void CLGenerateProposalsLayer::run()
+{
+    // Acquire all the temporaries
+    _memory_group.acquire();
+
+    // Compute all the anchors
+    CLScheduler::get().enqueue(_compute_anchors_kernel, false);
+
+    // Transpose and reshape the inputs
+    if(!_is_nhwc)
+    {
+        CLScheduler::get().enqueue(_permute_deltas_kernel, false);
+        CLScheduler::get().enqueue(_permute_scores_kernel, false);
+    }
+    CLScheduler::get().enqueue(_flatten_deltas_kernel, false);
+    CLScheduler::get().enqueue(_flatten_scores_kernel, false);
+
+    // Build the boxes
+    CLScheduler::get().enqueue(_bounding_box_kernel, false);
+    // Non maxima suppression
+    run_cpp_nms_kernel();
+    // Add dummy batch indexes
+    CLScheduler::get().enqueue(_memset_kernel, true);
+    CLScheduler::get().enqueue(_padded_copy_kernel, true);
+
+    // Release all the temporaries
+    _memory_group.release();
+}
+} // namespace arm_compute