COMPMID-1849: Implement CPPDetectionPostProcessLayer

* Add DetectionPostProcessLayer
* Add DetectionPostProcessLayer at the graph

Change-Id: I7e56f6cffc26f112d26dfe74853085bb8ec7d849
Signed-off-by: Isabella Gottardi <isabella.gottardi@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1639
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

9 files changed, 600 insertions, 12 deletions

diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index 54bd066712..228f2d211a 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -393,6 +393,36 @@ NodeID GraphBuilder::add_detection_output_node(Graph &g, NodeParams params, Node
     return detect_nid;
 }
 
+NodeID GraphBuilder::add_detection_post_process_node(Graph &g, NodeParams params, NodeIdxPair input_box_encoding, NodeIdxPair input_class_prediction, const DetectionPostProcessLayerInfo &detect_info,
+                                                     ITensorAccessorUPtr anchors_accessor, const QuantizationInfo &anchor_quant_info)
+{
+    check_nodeidx_pair(input_box_encoding, g);
+    check_nodeidx_pair(input_class_prediction, g);
+
+    // Get input tensor descriptor
+    const TensorDescriptor input_box_encoding_tensor_desc = get_tensor_descriptor(g, g.node(input_box_encoding.node_id)->outputs()[0]);
+
+    // Calculate anchor descriptor
+    TensorDescriptor anchor_desc = input_box_encoding_tensor_desc;
+    if(!anchor_quant_info.empty())
+    {
+        anchor_desc.quant_info = anchor_quant_info;
+    }
+
+    // Create anchors nodes
+    auto anchors_nid = add_const_node_with_name(g, params, "Anchors", anchor_desc, std::move(anchors_accessor));
+
+    // Create detection_output node and connect
+    NodeID detect_nid = g.add_node<DetectionPostProcessLayerNode>(detect_info);
+    g.add_connection(input_box_encoding.node_id, input_box_encoding.index, detect_nid, 0);
+    g.add_connection(input_class_prediction.node_id, input_class_prediction.index, detect_nid, 1);
+    g.add_connection(anchors_nid, 0, detect_nid, 2);
+
+    set_node_params(g, detect_nid, params);
+
+    return detect_nid;
+}
+
 NodeID GraphBuilder::add_dummy_node(Graph &g, NodeParams params, NodeIdxPair input, TensorShape shape)
 {
     return create_simple_single_input_output_node<DummyNode>(g, params, input, shape);
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index b9f22f6199..82b6dd6a54 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -166,6 +166,62 @@ std::unique_ptr<IFunction> create_detection_output_layer<CPPDetectionOutputLayer
 
     return std::move(wrap_function);
 }
+template <>
+std::unique_ptr<IFunction> create_detection_post_process_layer<CPPDetectionPostProcessLayer, CLTargetInfo>(DetectionPostProcessLayerNode &node)
+{
+    validate_node<CLTargetInfo>(node, 3 /* expected inputs */, 4 /* expected outputs */);
+
+    // Extract IO and info
+    CLTargetInfo::TensorType           *input0      = get_backing_tensor<CLTargetInfo>(node.input(0));
+    CLTargetInfo::TensorType           *input1      = get_backing_tensor<CLTargetInfo>(node.input(1));
+    CLTargetInfo::TensorType           *input2      = get_backing_tensor<CLTargetInfo>(node.input(2));
+    CLTargetInfo::TensorType           *output0     = get_backing_tensor<CLTargetInfo>(node.output(0));
+    CLTargetInfo::TensorType           *output1     = get_backing_tensor<CLTargetInfo>(node.output(1));
+    CLTargetInfo::TensorType           *output2     = get_backing_tensor<CLTargetInfo>(node.output(2));
+    CLTargetInfo::TensorType           *output3     = get_backing_tensor<CLTargetInfo>(node.output(3));
+    const DetectionPostProcessLayerInfo detect_info = node.detection_post_process_info();
+
+    ARM_COMPUTE_ERROR_ON(input0 == nullptr);
+    ARM_COMPUTE_ERROR_ON(input1 == nullptr);
+    ARM_COMPUTE_ERROR_ON(input2 == nullptr);
+    ARM_COMPUTE_ERROR_ON(output0 == nullptr);
+    ARM_COMPUTE_ERROR_ON(output1 == nullptr);
+    ARM_COMPUTE_ERROR_ON(output2 == nullptr);
+    ARM_COMPUTE_ERROR_ON(output3 == nullptr);
+
+    // Create and configure function
+    auto func = support::cpp14::make_unique<CPPDetectionPostProcessLayer>();
+    func->configure(input0, input1, input2, output0, output1, output2, output3, detect_info);
+
+    // Log info
+    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
+                               << node.name()
+                               << " Type: " << node.type()
+                               << " Target: " << CLTargetInfo::TargetType
+                               << " Data Type: " << input0->info()->data_type()
+                               << " Input0 shape: " << input0->info()->tensor_shape()
+                               << " Input1 shape: " << input1->info()->tensor_shape()
+                               << " Input2 shape: " << input2->info()->tensor_shape()
+                               << " Output0 shape: " << output0->info()->tensor_shape()
+                               << " Output1 shape: " << output1->info()->tensor_shape()
+                               << " Output2 shape: " << output2->info()->tensor_shape()
+                               << " Output3 shape: " << output3->info()->tensor_shape()
+                               << " DetectionPostProcessLayer info: " << detect_info
+                               << std::endl);
+
+    auto wrap_function = support::cpp14::make_unique<CPPWrapperFunction>();
+
+    wrap_function->register_function(std::move(func));
+    wrap_function->register_tensor(input0);
+    wrap_function->register_tensor(input1);
+    wrap_function->register_tensor(input2);
+    wrap_function->register_tensor(output0);
+    wrap_function->register_tensor(output1);
+    wrap_function->register_tensor(output2);
+    wrap_function->register_tensor(output3);
+
+    return std::move(wrap_function);
+}
 } // namespace detail
 
 std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &ctx)
@@ -196,6 +252,8 @@ std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &
             return detail::create_depthwise_convolution_layer<CLDepthwiseConvolutionLayerFunctions, CLTargetInfo>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::create_detection_output_layer<CPPDetectionOutputLayer, CLTargetInfo>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
+        case NodeType::DetectionPostProcessLayer:
+            return detail::create_detection_post_process_layer<CPPDetectionPostProcessLayer, CLTargetInfo>(*polymorphic_downcast<DetectionPostProcessLayerNode *>(node));
         case NodeType::EltwiseLayer:
             return detail::create_eltwise_layer<CLEltwiseFunctions, CLTargetInfo>(*polymorphic_downcast<EltwiseLayerNode *>(node));
         case NodeType::FlattenLayer:
diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp
index 78771102e8..40ec508767 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::DetectionPostProcessLayer:
+            return detail::validate_detection_post_process_layer<CPPDetectionPostProcessLayer>(*polymorphic_downcast<DetectionPostProcessLayerNode *>(node));
         case NodeType::GenerateProposalsLayer:
             return detail::validate_generate_proposals_layer<CLGenerateProposalsLayer>(*polymorphic_downcast<GenerateProposalsLayerNode *>(node));
         case NodeType::NormalizePlanarYUVLayer:
diff --git a/src/graph/backends/GLES/GCNodeValidator.cpp b/src/graph/backends/GLES/GCNodeValidator.cpp
index a767d7b107..9cbb9a12ef 100644
--- a/src/graph/backends/GLES/GCNodeValidator.cpp
+++ b/src/graph/backends/GLES/GCNodeValidator.cpp
@@ -113,6 +113,8 @@ Status GCNodeValidator::validate(INode *node)
             return validate_depthwise_convolution_layer(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : DetectionOutputLayer");
+        case NodeType::DetectionPostProcessLayer:
+            return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : DetectionPostProcessLayer");
         case NodeType::FlattenLayer:
             return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : FlattenLayer");
         case NodeType::GenerateProposalsLayer:
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index b808ef81f9..852de549fa 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -215,6 +215,8 @@ std::unique_ptr<IFunction> NEFunctionFactory::create(INode *node, GraphContext &
             return detail::create_depthwise_convolution_layer<NEDepthwiseConvolutionLayerFunctions, NETargetInfo>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::create_detection_output_layer<CPPDetectionOutputLayer, NETargetInfo>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
+        case NodeType::DetectionPostProcessLayer:
+            return detail::create_detection_post_process_layer<CPPDetectionPostProcessLayer, NETargetInfo>(*polymorphic_downcast<DetectionPostProcessLayerNode *>(node));
         case NodeType::EltwiseLayer:
             return detail::create_eltwise_layer<NEEltwiseFunctions, NETargetInfo>(*polymorphic_downcast<EltwiseLayerNode *>(node));
         case NodeType::FlattenLayer:
diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp
index 3b1d2aa59c..734b3401f7 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::DetectionPostProcessLayer:
+            return detail::validate_detection_post_process_layer<CPPDetectionPostProcessLayer>(*polymorphic_downcast<DetectionPostProcessLayerNode *>(node));
         case NodeType::GenerateProposalsLayer:
             return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : GenerateProposalsLayer");
         case NodeType::NormalizePlanarYUVLayer:
diff --git a/src/graph/nodes/DetectionPostProcessLayerNode.cpp b/src/graph/nodes/DetectionPostProcessLayerNode.cpp
new file mode 100644
index 0000000000..4a5df1ac4e
--- /dev/null
+++ b/src/graph/nodes/DetectionPostProcessLayerNode.cpp
@@ -0,0 +1,104 @@
+/*
+ * 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/DetectionPostProcessLayerNode.h"
+
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/graph/Graph.h"
+#include "arm_compute/graph/INodeVisitor.h"
+#include "arm_compute/graph/Utils.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+DetectionPostProcessLayerNode::DetectionPostProcessLayerNode(DetectionPostProcessLayerInfo detection_info)
+    : _info(detection_info)
+{
+    _input_edges.resize(3, EmptyEdgeID);
+    _outputs.resize(4, NullTensorID);
+}
+
+DetectionPostProcessLayerInfo DetectionPostProcessLayerNode::detection_post_process_info() const
+{
+    return _info;
+}
+
+bool DetectionPostProcessLayerNode::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) && (output_id(3) != NullTensorID))
+    {
+        for(unsigned int i = 0; i < 4; ++i)
+        {
+            Tensor *dst = output(i);
+            ARM_COMPUTE_ERROR_ON(dst == nullptr);
+            dst->desc() = configure_output(i);
+        }
+        return true;
+    }
+    return false;
+}
+
+TensorDescriptor DetectionPostProcessLayerNode::configure_output(size_t idx) const
+{
+    ARM_COMPUTE_UNUSED(idx);
+    ARM_COMPUTE_ERROR_ON(idx >= _outputs.size());
+
+    TensorDescriptor   output_desc;
+    const unsigned int num_detected_box = _info.max_detections() * _info.max_classes_per_detection();
+
+    switch(idx)
+    {
+        case 0:
+            // Configure boxes output
+            output_desc.shape = TensorShape(kNumCoordBox, num_detected_box, kBatchSize);
+            break;
+        case 1:
+        case 2:
+            // Configure classes or scores output
+            output_desc.shape = TensorShape(num_detected_box, kBatchSize);
+            break;
+        case 3:
+            // Configure num_detection
+            output_desc.shape = TensorShape(1);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported output index");
+    }
+    output_desc.data_type = DataType::F32;
+
+    return output_desc;
+}
+
+NodeType DetectionPostProcessLayerNode::type() const
+{
+    return NodeType::DetectionPostProcessLayer;
+}
+
+void DetectionPostProcessLayerNode::accept(INodeVisitor &v)
+{
+    v.visit(*this);
+}
+} // namespace graph
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp b/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp
index a1f4e6e89c..13a34b43cd 100644
--- a/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp
+++ b/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp
@@ -166,9 +166,9 @@ void retrieve_all_conf_scores(const ITensor *input_conf, const int num,
  * @param[out] all_location_predictions All the location predictions.
  *
  */
-void retrieve_all_priorbox(const ITensor               *input_priorbox,
-                           const int                    num_priors,
-                           std::vector<NormalizedBBox> &all_prior_bboxes,
+void retrieve_all_priorbox(const ITensor     *input_priorbox,
+                           const int          num_priors,
+                           std::vector<BBox> &all_prior_bboxes,
                            std::vector<std::array<float, 4>> &all_prior_variances)
 {
     for(int i = 0; i < num_priors; ++i)
@@ -206,9 +206,9 @@ void retrieve_all_priorbox(const ITensor               *input_priorbox,
  * @param[out] decode_bbox                The decoded bboxes.
  *
  */
-void DecodeBBox(const NormalizedBBox &prior_bbox, const std::array<float, 4> &prior_variance,
+void DecodeBBox(const BBox &prior_bbox, const std::array<float, 4> &prior_variance,
                 const DetectionOutputLayerCodeType code_type, const bool variance_encoded_in_target,
-                const bool clip_bbox, const NormalizedBBox &bbox, NormalizedBBox &decode_bbox)
+                const bool clip_bbox, const BBox &bbox, BBox &decode_bbox)
 {
     // if the variance is encoded in target, we simply need to add the offset predictions
     // otherwise we need to scale the offset accordingly.
@@ -287,7 +287,7 @@ void DecodeBBox(const NormalizedBBox &prior_bbox, const std::array<float, 4> &pr
  * @param[out] indices         The kept indices of bboxes after nms.
  *
  */
-void ApplyNMSFast(const std::vector<NormalizedBBox> &bboxes,
+void ApplyNMSFast(const std::vector<BBox> &bboxes,
                   const std::vector<float> &scores, const float score_threshold,
                   const float nms_threshold, const float eta, const int top_k,
                   std::vector<int> &indices)
@@ -329,7 +329,7 @@ void ApplyNMSFast(const std::vector<NormalizedBBox> &bboxes,
             if(keep)
             {
                 // Compute the jaccard (intersection over union IoU) overlap between two bboxes.
-                NormalizedBBox intersect_bbox = std::array<float, 4>({ { 0, 0, 0, 0 } });
+                BBox intersect_bbox = std::array<float, 4>({ 0, 0, 0, 0 });
                 if(bboxes[kept_idx][0] > bboxes[idx][2] || bboxes[kept_idx][2] < bboxes[idx][0] || bboxes[kept_idx][1] > bboxes[idx][3] || bboxes[kept_idx][3] < bboxes[idx][1])
                 {
                     intersect_bbox = std::array<float, 4>({ { 0, 0, 0, 0 } });
@@ -466,7 +466,7 @@ void CPPDetectionOutputLayer::run()
             }
             ARM_COMPUTE_ERROR_ON_MSG(_all_location_predictions[i].find(label) == _all_location_predictions[i].end(), "Could not find location predictions for label %d.", label);
 
-            const std::vector<NormalizedBBox> &label_loc_preds = _all_location_predictions[i].find(label)->second;
+            const std::vector<BBox> &label_loc_preds = _all_location_predictions[i].find(label)->second;
 
             const int num_bboxes = _all_prior_bboxes.size();
             ARM_COMPUTE_ERROR_ON(_all_prior_variances[i].size() != 4);
@@ -499,8 +499,8 @@ void CPPDetectionOutputLayer::run()
             {
                 ARM_COMPUTE_ERROR("Could not find predictions for label %d.", label);
             }
-            const std::vector<float>          &scores = conf_scores.find(c)->second;
-            const std::vector<NormalizedBBox> &bboxes = decode_bboxes.find(label)->second;
+            const std::vector<float> &scores = conf_scores.find(c)->second;
+            const std::vector<BBox> &bboxes = decode_bboxes.find(label)->second;
 
             ApplyNMSFast(bboxes, scores, _info.confidence_threshold(), _info.nms_threshold(), _info.eta(), _info.top_k(), indices[c]);
 
@@ -572,8 +572,8 @@ void CPPDetectionOutputLayer::run()
                 // or there are no location predictions for current label.
                 ARM_COMPUTE_ERROR("Could not find predictions for the label %d.", label);
             }
-            const std::vector<NormalizedBBox> &bboxes  = decode_bboxes.find(loc_label)->second;
-            const std::vector<int>            &indices = it.second;
+            const std::vector<BBox> &bboxes  = decode_bboxes.find(loc_label)->second;
+            const std::vector<int> &indices = it.second;
 
             for(auto idx : indices)
             {
diff --git a/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp b/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp
new file mode 100644
index 0000000000..2997b593c6
--- /dev/null
+++ b/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp
@@ -0,0 +1,388 @@
+/*
+ * 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/CPP/functions/CPPDetectionPostProcessLayer.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Validate.h"
+#include "support/ToolchainSupport.h"
+
+#include <cstddef>
+#include <ios>
+#include <list>
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input_box_encoding, const ITensorInfo *input_class_score, const ITensorInfo *input_anchors,
+                          ITensorInfo *output_boxes, ITensorInfo *output_classes, ITensorInfo *output_scores, ITensorInfo *num_detection,
+                          DetectionPostProcessLayerInfo info, const unsigned int kBatchSize, const unsigned int kNumCoordBox)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input_box_encoding, input_class_score, input_anchors);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_box_encoding, 1, DataType::F32, DataType::QASYMM8);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_box_encoding, input_class_score, input_anchors);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_box_encoding->num_dimensions() > 3, "The location input tensor shape should be [4, N, kBatchSize].");
+    if(input_box_encoding->num_dimensions() > 2)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_box_encoding->dimension(2) != kBatchSize, "The third dimension of the input box_encoding tensor should be equal to %d.", kBatchSize);
+    }
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_box_encoding->dimension(0) != kNumCoordBox, "The first dimension of the input box_encoding tensor should be equal to %d.", kNumCoordBox);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_class_score->dimension(0) != (info.num_classes() + 1),
+                                    "The first dimension of the input class_prediction should be equal to the number of classes plus one.");
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_anchors->num_dimensions() > 3, "The anchors input tensor shape should be [4, N, kBatchSize].");
+    if(input_anchors->num_dimensions() > 2)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_anchors->dimension(0) != kNumCoordBox, "The first dimension of the input anchors tensor should be equal to %d.", kNumCoordBox);
+    }
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((input_box_encoding->dimension(1) != input_class_score->dimension(1))
+                                    || (input_box_encoding->dimension(1) != input_anchors->dimension(1)),
+                                    "The second dimension of the inputs should be the same.");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_detection->num_dimensions() > 1, "The num_detection output tensor shape should be [M].");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((info.iou_threshold() <= 0.0f) || (info.iou_threshold() > 1.0f), "The intersection over union should be positive and less than 1.");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.max_classes_per_detection() <= 0, "The number of max classes per detection should be positive.");
+
+    const unsigned int num_detected_boxes = info.max_detections() * info.max_classes_per_detection();
+
+    // Validate configured outputs
+    if(output_boxes->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output_boxes->tensor_shape(), TensorShape(4U, num_detected_boxes, 1U));
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_boxes, 1, DataType::F32);
+    }
+    if(output_classes->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output_classes->tensor_shape(), TensorShape(num_detected_boxes, 1U));
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_classes, 1, DataType::F32);
+    }
+    if(output_scores->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output_scores->tensor_shape(), TensorShape(num_detected_boxes, 1U));
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_scores, 1, DataType::F32);
+    }
+    if(num_detection->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(num_detection->tensor_shape(), TensorShape(1U));
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(num_detection, 1, DataType::F32);
+    }
+
+    return Status{};
+}
+
+/** Decode a bbox according to a anchors and scale info.
+ *
+ * @param[in]  input_box_encoding The input prior bounding boxes.
+ * @param[in]  input_anchors      The corresponding input variance.
+ * @param[in]  info               The detection informations
+ * @param[out] decoded_boxes      The decoded bboxes.
+ */
+void DecodeCenterSizeBoxes(const ITensor *input_box_encoding, const ITensor *input_anchors, DetectionPostProcessLayerInfo info, Tensor *decoded_boxes)
+{
+    const QuantizationInfo &qi_box     = input_box_encoding->info()->quantization_info();
+    const QuantizationInfo &qi_anchors = input_anchors->info()->quantization_info();
+    BBox                    box_centersize;
+    BBox                    anchor;
+
+    Window win;
+    win.use_tensor_dimensions(input_box_encoding->info()->tensor_shape());
+    win.set_dimension_step(0U, 4U);
+    win.set_dimension_step(1U, 1U);
+    Iterator box_it(input_box_encoding, win);
+    Iterator anchor_it(input_anchors, win);
+    Iterator decoded_it(decoded_boxes, win);
+
+    const float half_factor = 0.5f;
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        if(is_data_type_quantized(input_box_encoding->info()->data_type()))
+        {
+            const auto box_ptr    = reinterpret_cast<const qasymm8_t *>(box_it.ptr());
+            const auto anchor_ptr = reinterpret_cast<const qasymm8_t *>(anchor_it.ptr());
+            box_centersize        = BBox({ dequantize_qasymm8(*box_ptr, qi_box), dequantize_qasymm8(*(box_ptr + 1), qi_box),
+                                           dequantize_qasymm8(*(2 + box_ptr), qi_box), dequantize_qasymm8(*(3 + box_ptr), qi_box)
+                                         });
+            anchor = BBox({ dequantize_qasymm8(*anchor_ptr, qi_anchors), dequantize_qasymm8(*(anchor_ptr + 1), qi_anchors),
+                            dequantize_qasymm8(*(2 + anchor_ptr), qi_anchors), dequantize_qasymm8(*(3 + anchor_ptr), qi_anchors)
+                          });
+        }
+        else
+        {
+            const auto box_ptr    = reinterpret_cast<const float *>(box_it.ptr());
+            const auto anchor_ptr = reinterpret_cast<const float *>(anchor_it.ptr());
+            box_centersize        = BBox({ *box_ptr, *(box_ptr + 1), *(2 + box_ptr), *(3 + box_ptr) });
+            anchor                = BBox({ *anchor_ptr, *(anchor_ptr + 1), *(2 + anchor_ptr), *(3 + anchor_ptr) });
+        }
+
+        // BBox is equavalent to CenterSizeEncoding [y,x,h,w]
+        const float y_center = box_centersize[0] / info.scale_value_y() * anchor[2] + anchor[0];
+        const float x_center = box_centersize[1] / info.scale_value_x() * anchor[3] + anchor[1];
+        const float half_h   = half_factor * static_cast<float>(std::exp(box_centersize[2] / info.scale_value_h())) * anchor[2];
+        const float half_w   = half_factor * static_cast<float>(std::exp(box_centersize[3] / info.scale_value_w())) * anchor[3];
+
+        // Box Corner encoding boxes are saved as [xmin, ymin, xmax, ymax]
+        auto decoded_ptr   = reinterpret_cast<float *>(decoded_it.ptr());
+        *(decoded_ptr)     = x_center - half_w; // xmin
+        *(1 + decoded_ptr) = y_center - half_h; // ymin
+        *(2 + decoded_ptr) = x_center + half_w; // xmax
+        *(3 + decoded_ptr) = y_center + half_h; // ymax
+    },
+    box_it, anchor_it, decoded_it);
+}
+
+void SaveOutputs(const Tensor *decoded_boxes, const std::vector<int> &result_idx_boxes_after_nms, const std::vector<float> &result_scores_after_nms, const std::vector<int> &result_classes_after_nms,
+                 std::vector<unsigned int> &sorted_indices, const unsigned int num_output, const unsigned int max_detections, ITensor *output_boxes, ITensor *output_classes, ITensor *output_scores,
+                 ITensor *num_detection)
+{
+    // ymin,xmin,ymax,xmax -> xmin,ymin,xmax,ymax
+    unsigned int i = 0;
+    for(; i < num_output; ++i)
+    {
+        const unsigned int box_in_idx = result_idx_boxes_after_nms[sorted_indices[i]];
+        *(reinterpret_cast<float *>(output_boxes->ptr_to_element(Coordinates(0, i)))) = *(reinterpret_cast<float *>(decoded_boxes->ptr_to_element(Coordinates(1, box_in_idx))));
+        *(reinterpret_cast<float *>(output_boxes->ptr_to_element(Coordinates(1, i)))) = *(reinterpret_cast<float *>(decoded_boxes->ptr_to_element(Coordinates(0, box_in_idx))));
+        *(reinterpret_cast<float *>(output_boxes->ptr_to_element(Coordinates(2, i)))) = *(reinterpret_cast<float *>(decoded_boxes->ptr_to_element(Coordinates(3, box_in_idx))));
+        *(reinterpret_cast<float *>(output_boxes->ptr_to_element(Coordinates(3, i)))) = *(reinterpret_cast<float *>(decoded_boxes->ptr_to_element(Coordinates(2, box_in_idx))));
+        *(reinterpret_cast<float *>(output_classes->ptr_to_element(Coordinates(i)))) = static_cast<float>(result_classes_after_nms[sorted_indices[i]]);
+        *(reinterpret_cast<float *>(output_scores->ptr_to_element(Coordinates(i))))  = result_scores_after_nms[sorted_indices[i]];
+    }
+    for(; i < max_detections; ++i)
+    {
+        *(reinterpret_cast<float *>(output_boxes->ptr_to_element(Coordinates(1, i)))) = 0.0f;
+        *(reinterpret_cast<float *>(output_boxes->ptr_to_element(Coordinates(0, i)))) = 0.0f;
+        *(reinterpret_cast<float *>(output_boxes->ptr_to_element(Coordinates(3, i)))) = 0.0f;
+        *(reinterpret_cast<float *>(output_boxes->ptr_to_element(Coordinates(2, i)))) = 0.0f;
+        *(reinterpret_cast<float *>(output_classes->ptr_to_element(Coordinates(i)))) = 0.0f;
+        *(reinterpret_cast<float *>(output_scores->ptr_to_element(Coordinates(i))))  = 0.0f;
+    }
+    *(reinterpret_cast<float *>(num_detection->ptr_to_element(Coordinates(0)))) = num_output;
+}
+} // namespace
+
+CPPDetectionPostProcessLayer::CPPDetectionPostProcessLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _nms(), _input_box_encoding(nullptr), _input_scores(nullptr), _input_anchors(nullptr), _output_boxes(nullptr), _output_classes(nullptr),
+      _output_scores(nullptr), _num_detection(nullptr), _info(), _num_boxes(), _num_classes_with_background(), _num_max_detected_boxes(), _decoded_boxes(), _decoded_scores(), _selected_indices(),
+      _class_scores(), _input_scores_to_use(nullptr), _result_idx_boxes_after_nms(), _result_classes_after_nms(), _result_scores_after_nms(), _sorted_indices(), _box_scores()
+{
+}
+
+void CPPDetectionPostProcessLayer::configure(const ITensor *input_box_encoding, const ITensor *input_scores, const ITensor *input_anchors,
+                                             ITensor *output_boxes, ITensor *output_classes, ITensor *output_scores, ITensor *num_detection, DetectionPostProcessLayerInfo info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input_box_encoding, input_scores, input_anchors, output_boxes, output_classes, output_scores);
+    _num_max_detected_boxes = info.max_detections() * info.max_classes_per_detection();
+
+    auto_init_if_empty(*output_boxes->info(), TensorInfo(TensorShape(_kNumCoordBox, _num_max_detected_boxes, _kBatchSize), 1, DataType::F32));
+    auto_init_if_empty(*output_classes->info(), TensorInfo(TensorShape(_num_max_detected_boxes, _kBatchSize), 1, DataType::F32));
+    auto_init_if_empty(*output_scores->info(), TensorInfo(TensorShape(_num_max_detected_boxes, _kBatchSize), 1, DataType::F32));
+    auto_init_if_empty(*num_detection->info(), TensorInfo(TensorShape(1U), 1, DataType::F32));
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input_box_encoding->info(), input_scores->info(), input_anchors->info(), output_boxes->info(), output_classes->info(), output_scores->info(),
+                                                  num_detection->info(),
+                                                  info, _kBatchSize, _kNumCoordBox));
+
+    _input_box_encoding          = input_box_encoding;
+    _input_scores                = input_scores;
+    _input_anchors               = input_anchors;
+    _output_boxes                = output_boxes;
+    _output_classes              = output_classes;
+    _output_scores               = output_scores;
+    _num_detection               = num_detection;
+    _info                        = info;
+    _num_boxes                   = input_box_encoding->info()->dimension(1);
+    _num_classes_with_background = _input_scores->info()->dimension(0);
+
+    auto_init_if_empty(*_decoded_boxes.info(), TensorInfo(TensorShape(_kNumCoordBox, _input_box_encoding->info()->dimension(1), _kBatchSize), 1, DataType::F32));
+    auto_init_if_empty(*_decoded_scores.info(), TensorInfo(TensorShape(_input_scores->info()->dimension(0), _input_scores->info()->dimension(1), _kBatchSize), 1, DataType::F32));
+    auto_init_if_empty(*_selected_indices.info(), TensorInfo(TensorShape(info.max_detections()), 1, DataType::S32));
+
+    const unsigned int num_classes_per_box = std::min(info.max_classes_per_detection(), info.num_classes());
+    auto_init_if_empty(*_class_scores.info(), TensorInfo(info.use_regular_nms() ? TensorShape(_num_boxes) : TensorShape(_num_boxes * num_classes_per_box), 1, DataType::F32));
+
+    _input_scores_to_use = is_data_type_quantized(input_box_encoding->info()->data_type()) ? &_decoded_scores : _input_scores;
+
+    // Manage intermediate buffers
+    _memory_group.manage(&_decoded_boxes);
+    _memory_group.manage(&_decoded_scores);
+    _memory_group.manage(&_selected_indices);
+    _memory_group.manage(&_class_scores);
+    _nms.configure(&_decoded_boxes, &_class_scores, &_selected_indices, info.use_regular_nms() ? info.detection_per_class() : info.max_detections(), info.nms_score_threshold(), info.iou_threshold());
+
+    // Allocate and reserve intermediate tensors and vectors
+    _decoded_boxes.allocator()->allocate();
+    _decoded_scores.allocator()->allocate();
+    _selected_indices.allocator()->allocate();
+    _class_scores.allocator()->allocate();
+
+    if(info.use_regular_nms())
+    {
+        _result_idx_boxes_after_nms.reserve(_info.detection_per_class() * _info.num_classes());
+        _result_classes_after_nms.reserve(_info.detection_per_class() * _info.num_classes());
+        _result_scores_after_nms.reserve(_info.detection_per_class() * _info.num_classes());
+    }
+    else
+    {
+        _result_scores_after_nms.reserve(num_classes_per_box * _num_boxes);
+        _result_classes_after_nms.reserve(num_classes_per_box * _num_boxes);
+        _result_scores_after_nms.reserve(num_classes_per_box * _num_boxes);
+        _box_scores.reserve(_num_boxes);
+    }
+    _sorted_indices.resize(info.use_regular_nms() ? info.max_detections() : info.num_classes());
+}
+
+Status CPPDetectionPostProcessLayer::validate(const ITensorInfo *input_box_encoding, const ITensorInfo *input_class_score, const ITensorInfo *input_anchors,
+                                              ITensorInfo *output_boxes, ITensorInfo *output_classes, ITensorInfo *output_scores, ITensorInfo *num_detection, DetectionPostProcessLayerInfo info)
+{
+    constexpr unsigned int kBatchSize             = 1;
+    constexpr unsigned int kNumCoordBox           = 4;
+    const TensorInfo       _decoded_boxes_info    = TensorInfo(TensorShape(kNumCoordBox, input_box_encoding->dimension(1)), 1, DataType::F32);
+    const TensorInfo       _decoded_scores_info   = TensorInfo(TensorShape(input_box_encoding->dimension(1)), 1, DataType::F32);
+    const TensorInfo       _selected_indices_info = TensorInfo(TensorShape(info.max_detections()), 1, DataType::S32);
+
+    ARM_COMPUTE_RETURN_ON_ERROR(CPPNonMaximumSuppression::validate(&_decoded_boxes_info, &_decoded_scores_info, &_selected_indices_info, info.max_detections(), info.nms_score_threshold(),
+                                                                   info.iou_threshold()));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input_box_encoding, input_class_score, input_anchors, output_boxes, output_classes, output_scores, num_detection, info, kBatchSize, kNumCoordBox));
+
+    return Status{};
+}
+
+void CPPDetectionPostProcessLayer::run()
+{
+    const unsigned int num_classes    = _info.num_classes();
+    const unsigned int max_detections = _info.max_detections();
+
+    DecodeCenterSizeBoxes(_input_box_encoding, _input_anchors, _info, &_decoded_boxes);
+
+    // Decode scores if necessary
+    if(is_data_type_quantized(_input_box_encoding->info()->data_type()))
+    {
+        for(unsigned int idx_c = 0; idx_c < _num_classes_with_background; ++idx_c)
+        {
+            for(unsigned int idx_b = 0; idx_b < _num_boxes; ++idx_b)
+            {
+                *(reinterpret_cast<float *>(_decoded_scores.ptr_to_element(Coordinates(idx_c, idx_b)))) =
+                    dequantize_qasymm8(*(reinterpret_cast<qasymm8_t *>(_input_scores->ptr_to_element(Coordinates(idx_c, idx_b)))), _input_scores->info()->quantization_info());
+            }
+        }
+    }
+    // Regular NMS
+    if(_info.use_regular_nms())
+    {
+        for(unsigned int c = 0; c < num_classes; ++c)
+        {
+            // For each boxes get scores of the boxes for the class c
+            for(unsigned int i = 0; i < _num_boxes; ++i)
+            {
+                *(reinterpret_cast<float *>(_class_scores.ptr_to_element(Coordinates(i)))) =
+                    *(reinterpret_cast<float *>(_input_scores_to_use->ptr_to_element(Coordinates(c + 1, i)))); // i * _num_classes_with_background + c + 1
+            }
+            _nms.run();
+
+            for(unsigned int i = 0; i < _info.detection_per_class(); ++i)
+            {
+                const auto selected_index = *(reinterpret_cast<int *>(_selected_indices.ptr_to_element(Coordinates(i))));
+                if(selected_index == -1)
+                {
+                    // Nms will return -1 for all the last M-elements not valid
+                    continue;
+                }
+                _result_idx_boxes_after_nms.emplace_back(selected_index);
+                _result_scores_after_nms.emplace_back((reinterpret_cast<float *>(_class_scores.buffer()))[selected_index]);
+                _result_classes_after_nms.emplace_back(c);
+            }
+        }
+
+        // We select the max detection numbers of the highest score of all classes
+        const auto num_selected = _result_idx_boxes_after_nms.size();
+        const auto num_output   = std::min<unsigned int>(max_detections, num_selected);
+
+        // Sort selected indices based on result scores
+        std::iota(_sorted_indices.begin(), _sorted_indices.end(), 0);
+        std::partial_sort(_sorted_indices.data(),
+                          _sorted_indices.data() + num_output,
+                          _sorted_indices.data() + num_selected,
+                          [&](unsigned int first, unsigned int second)
+        {
+
+            return _result_scores_after_nms[first] > _result_scores_after_nms[second];
+        });
+
+        SaveOutputs(&_decoded_boxes, _result_idx_boxes_after_nms, _result_scores_after_nms, _result_classes_after_nms,
+                    _sorted_indices, num_output, max_detections, _output_boxes, _output_classes, _output_scores, _num_detection);
+    }
+    // Fast NMS
+    else
+    {
+        const unsigned int num_classes_per_box = std::min<unsigned int>(_info.max_classes_per_detection(), _info.num_classes());
+        for(unsigned int b = 0, index = 0; b < _num_boxes; ++b)
+        {
+            _box_scores.clear();
+            _sorted_indices.clear();
+            for(unsigned int c = 0; c < num_classes; ++c)
+            {
+                _box_scores.emplace_back(*(reinterpret_cast<float *>(_input_scores_to_use->ptr_to_element(Coordinates(c + 1, b)))));
+                _sorted_indices.push_back(c);
+            }
+            std::partial_sort(_sorted_indices.data(),
+                              _sorted_indices.data() + num_classes_per_box,
+                              _sorted_indices.data() + num_classes,
+                              [&](unsigned int first, unsigned int second)
+            {
+                return _box_scores[first] > _box_scores[second];
+            });
+
+            for(unsigned int i = 0; i < num_classes_per_box; ++i, ++index)
+            {
+                const float score_to_add                                                       = _box_scores[_sorted_indices[i]];
+                *(reinterpret_cast<float *>(_class_scores.ptr_to_element(Coordinates(index)))) = score_to_add;
+                _result_scores_after_nms.emplace_back(score_to_add);
+                _result_idx_boxes_after_nms.emplace_back(b);
+                _result_classes_after_nms.emplace_back(_sorted_indices[i]);
+            }
+        }
+
+        // Run NMS
+        _nms.run();
+
+        _sorted_indices.clear();
+        for(unsigned int i = 0; i < max_detections; ++i)
+        {
+            // NMS returns M valid indices, the not valid tail is filled with -1
+            if(*(reinterpret_cast<int *>(_selected_indices.ptr_to_element(Coordinates(i)))) == -1)
+            {
+                // Nms will return -1 for all the last M-elements not valid
+                break;
+            }
+            _sorted_indices.emplace_back(*(reinterpret_cast<int *>(_selected_indices.ptr_to_element(Coordinates(i)))));
+        }
+        // We select the max detection numbers of the highest score of all classes
+        const auto num_output = std::min<unsigned int>(_info.max_detections(), _sorted_indices.size());
+
+        SaveOutputs(&_decoded_boxes, _result_idx_boxes_after_nms, _result_scores_after_nms, _result_classes_after_nms,
+                    _sorted_indices, num_output, max_detections, _output_boxes, _output_classes, _output_scores, _num_detection);
+    }
+}
+} // namespace arm_compute
\ No newline at end of file