Face detection demo from Emza Visual Sense
Signed-off-by: Michael Levit michaell@emza-vs.com
Change-Id: I7958b05b5dbe9a785e0f8a241b716c17a9ca976f
diff --git a/source/use_case/object_detection/include/DetectionResult.hpp b/source/use_case/object_detection/include/DetectionResult.hpp
new file mode 100644
index 0000000..78895f7
--- /dev/null
+++ b/source/use_case/object_detection/include/DetectionResult.hpp
@@ -0,0 +1,51 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef DETECTION_RESULT_HPP
+#define DETECTION_RESULT_HPP
+
+
+namespace arm {
+namespace app {
+
+ /**
+ * @brief Class representing a single detection result.
+ */
+ class DetectionResult {
+ public:
+ double m_normalisedVal{0.0};
+ int m_x0{0};
+ int m_y0{0};
+ int m_w{0};
+ int m_h{0};
+
+ DetectionResult() = default;
+ ~DetectionResult() = default;
+
+ DetectionResult(double normalisedVal,int x0,int y0, int w,int h) :
+ m_normalisedVal(normalisedVal),
+ m_x0(x0),
+ m_y0(y0),
+ m_w(w),
+ m_h(h)
+ {
+ }
+ };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* DETECTION_RESULT_HPP */
diff --git a/source/use_case/object_detection/include/DetectionUseCaseUtils.hpp b/source/use_case/object_detection/include/DetectionUseCaseUtils.hpp
new file mode 100644
index 0000000..8ef48ac
--- /dev/null
+++ b/source/use_case/object_detection/include/DetectionUseCaseUtils.hpp
@@ -0,0 +1,72 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef DETECTION_USE_CASE_UTILS_HPP
+#define DETECTION_USE_CASE_UTILS_HPP
+
+#include "hal.h"
+#include "DetectionResult.hpp"
+#include "UseCaseHandler.hpp" /* Handlers for different user options. */
+#include <inttypes.h>
+#include <vector>
+
+
+void DisplayDetectionMenu();
+
+namespace image{
+
+
+ /**
+ * @brief Presents inference results using the data presentation
+ * object.
+ * @param[in] platform Reference to the hal platform object.
+ * @param[in] results Vector of detection results to be displayed.
+ * @return true if successful, false otherwise.
+ **/
+ bool PresentInferenceResult(hal_platform & platform,
+ const std::vector < arm::app::DetectionResult > & results);
+
+
+ /**
+ * @brief Presents inference results along with the inference time using the data presentation
+ * object.
+ * @param[in] platform Reference to the hal platform object.
+ * @param[in] results Vector of detection results to be displayed.
+ * @param[in] infTimeMs Inference time in ms.
+ * @return true if successful, false otherwise.
+ **/
+ bool PresentInferenceResult(hal_platform & platform,
+ const std::vector < arm::app::DetectionResult > & results,
+ const time_t infTimeMs);
+
+ /**
+ * @brief Presents inference results along with the inference time using the data presentation
+ * object.
+ * @param[in] platform Reference to the hal platform object.
+ * @param[in] results Vector of detection results to be displayed.
+ * @param[in] infTimeMs Inference time in ms.
+ * @return true if successful, false otherwise.
+ **/
+ bool PresentInferenceResult(hal_platform & platform,
+ const std::vector < arm::app::DetectionResult > & results,
+ bool profilingEnabled,
+ const time_t infTimeMs = 0);
+ }
+
+
+
+
+#endif /* DETECTION_USE_CASE_UTILS_HPP */
diff --git a/source/use_case/object_detection/include/DetectorPostProcessing.hpp b/source/use_case/object_detection/include/DetectorPostProcessing.hpp
new file mode 100644
index 0000000..9a8549c
--- /dev/null
+++ b/source/use_case/object_detection/include/DetectorPostProcessing.hpp
@@ -0,0 +1,55 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef DETECTOR_POST_PROCESSING_HPP
+#define DETECTOR_POST_PROCESSING_HPP
+
+#include "UseCaseCommonUtils.hpp"
+#include "DetectionResult.hpp"
+
+namespace arm {
+namespace app {
+
+#if DISPLAY_RGB_IMAGE
+#define FORMAT_MULTIPLY_FACTOR 3
+#else
+#define FORMAT_MULTIPLY_FACTOR 1
+#endif /* DISPLAY_RGB_IMAGE */
+
+ /**
+ * @brief Post processing part of Yolo object detection CNN
+ * @param[in] img_in Pointer to the input image,detection bounding boxes drown on it.
+ * @param[in] model_output Output tesnsors after CNN invoked
+ * @param[out] results_out Vector of detected results.
+ * @return void
+ **/
+void RunPostProcessing(uint8_t *img_in,TfLiteTensor* model_output[2],std::vector<arm::app::DetectionResult> & results_out);
+
+
+ /**
+ * @brief Converts RGB image to grayscale
+ * @param[in] rgb Pointer to RGB input image
+ * @param[out] gray Pointer to RGB out image
+ * @param[in] im_w Input image width
+ * @param[in] im_h Input image height
+ * @return void
+ **/
+void RgbToGrayscale(const uint8_t *rgb,uint8_t *gray, int im_w,int im_h);
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* DETECTOR_POST_PROCESSING_HPP */
diff --git a/source/use_case/object_detection/include/UseCaseHandler.hpp b/source/use_case/object_detection/include/UseCaseHandler.hpp
new file mode 100644
index 0000000..56629c8
--- /dev/null
+++ b/source/use_case/object_detection/include/UseCaseHandler.hpp
@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef OBJ_DET_HANDLER_HPP
+#define OBJ_DET_HANDLER_HPP
+
+#include "AppContext.hpp"
+
+namespace arm {
+namespace app {
+
+ /**
+ * @brief Handles the inference event.
+ * @param[in] ctx Pointer to the application context.
+ * @param[in] imgIndex Index to the image to run object detection.
+ * @param[in] runAll Flag to request classification of all the available images.
+ * @return true or false based on execution success.
+ **/
+ bool ObjectDetectionHandler(ApplicationContext& ctx, uint32_t imgIndex, bool runAll);
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* OBJ_DET_HANDLER_HPP */
diff --git a/source/use_case/object_detection/include/YoloFastestModel.hpp b/source/use_case/object_detection/include/YoloFastestModel.hpp
new file mode 100644
index 0000000..f5709ea
--- /dev/null
+++ b/source/use_case/object_detection/include/YoloFastestModel.hpp
@@ -0,0 +1,55 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef YOLO_FASTEST_MODEL_HPP
+#define YOLO_FASTEST_MODEL_HPP
+
+#include "Model.hpp"
+
+namespace arm {
+namespace app {
+
+ class YoloFastestModel : public Model {
+
+ public:
+ /* Indices for the expected model - based on input tensor shape */
+ static constexpr uint32_t ms_inputRowsIdx = 1;
+ static constexpr uint32_t ms_inputColsIdx = 2;
+ static constexpr uint32_t ms_inputChannelsIdx = 3;
+
+ protected:
+ /** @brief Gets the reference to op resolver interface class. */
+ const tflite::MicroOpResolver& GetOpResolver() override;
+
+ /** @brief Adds operations to the op resolver instance. */
+ bool EnlistOperations() override;
+
+ const uint8_t* ModelPointer() override;
+
+ size_t ModelSize() override;
+
+ private:
+ /* Maximum number of individual operations that can be enlisted. */
+ static constexpr int ms_maxOpCnt = 8;
+
+ /* A mutable op resolver instance. */
+ tflite::MicroMutableOpResolver<ms_maxOpCnt> m_opResolver;
+ };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* YOLO_FASTEST_MODEL_HPP */
diff --git a/source/use_case/object_detection/src/DetectionUseCaseUtils.cc b/source/use_case/object_detection/src/DetectionUseCaseUtils.cc
new file mode 100644
index 0000000..1713c7e
--- /dev/null
+++ b/source/use_case/object_detection/src/DetectionUseCaseUtils.cc
@@ -0,0 +1,105 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "DetectionUseCaseUtils.hpp"
+#include "UseCaseCommonUtils.hpp"
+#include "InputFiles.hpp"
+#include <inttypes.h>
+
+
+void DisplayDetectionMenu()
+{
+ printf("\n\n");
+ printf("User input required\n");
+ printf("Enter option number from:\n\n");
+ printf(" %u. Run detection on next ifm\n", common::MENU_OPT_RUN_INF_NEXT);
+ printf(" %u. Run detection ifm at chosen index\n", common::MENU_OPT_RUN_INF_CHOSEN);
+ printf(" %u. Run detection on all ifm\n", common::MENU_OPT_RUN_INF_ALL);
+ printf(" %u. Show NN model info\n", common::MENU_OPT_SHOW_MODEL_INFO);
+ printf(" %u. List ifm\n\n", common::MENU_OPT_LIST_IFM);
+ printf(" Choice: ");
+ fflush(stdout);
+}
+
+
+bool image::PresentInferenceResult(hal_platform& platform,
+ const std::vector<arm::app::DetectionResult>& results)
+{
+ return PresentInferenceResult(platform, results, false);
+}
+
+bool image::PresentInferenceResult(hal_platform &platform,
+ const std::vector<arm::app::DetectionResult> &results,
+ const time_t infTimeMs)
+{
+ return PresentInferenceResult(platform, results, true, infTimeMs);
+}
+
+
+bool image::PresentInferenceResult(hal_platform &platform,
+ const std::vector<arm::app::DetectionResult> &results,
+ bool profilingEnabled,
+ const time_t infTimeMs)
+{
+ constexpr uint32_t dataPsnTxtStartX1 = 150;
+ constexpr uint32_t dataPsnTxtStartY1 = 30;
+
+
+ if(profilingEnabled)
+ {
+ platform.data_psn->set_text_color(COLOR_YELLOW);
+
+ /* If profiling is enabled, and the time is valid. */
+ info("Final results:\n");
+ info("Total number of inferences: 1\n");
+ if (infTimeMs)
+ {
+ std::string strInf =
+ std::string{"Inference: "} +
+ std::to_string(infTimeMs) +
+ std::string{"ms"};
+ platform.data_psn->present_data_text(
+ strInf.c_str(), strInf.size(),
+ dataPsnTxtStartX1, dataPsnTxtStartY1, 0);
+ }
+ }
+ platform.data_psn->set_text_color(COLOR_GREEN);
+
+ if(!profilingEnabled) {
+ info("Final results:\n");
+ info("Total number of inferences: 1\n");
+ }
+
+ for (uint32_t i = 0; i < results.size(); ++i) {
+
+ if(profilingEnabled) {
+ info("%" PRIu32 ") (%f) -> %s {x=%d,y=%d,w=%d,h=%d}\n", i,
+ results[i].m_normalisedVal, "Detection box:",
+ results[i].m_x0, results[i].m_y0, results[i].m_w, results[i].m_h );
+ }
+ else
+ {
+ info("%" PRIu32 ") (%f) -> %s {x=%d,y=%d,w=%d,h=%d}\n", i,
+ results[i].m_normalisedVal, "Detection box:",
+ results[i].m_x0, results[i].m_y0, results[i].m_w, results[i].m_h );
+ }
+ }
+
+ return true;
+}
+
+
+
diff --git a/source/use_case/object_detection/src/DetectorPostProcessing.cc b/source/use_case/object_detection/src/DetectorPostProcessing.cc
new file mode 100755
index 0000000..e781b62
--- /dev/null
+++ b/source/use_case/object_detection/src/DetectorPostProcessing.cc
@@ -0,0 +1,447 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "DetectorPostProcessing.hpp"
+#include <algorithm>
+#include <cmath>
+#include <stdint.h>
+#include <forward_list>
+
+
+typedef struct boxabs {
+ float left, right, top, bot;
+} boxabs;
+
+
+typedef struct branch {
+ int resolution;
+ int num_box;
+ float *anchor;
+ int8_t *tf_output;
+ float scale;
+ int zero_point;
+ size_t size;
+ float scale_x_y;
+} branch;
+
+typedef struct network {
+ int input_w;
+ int input_h;
+ int num_classes;
+ int num_branch;
+ branch *branchs;
+ int topN;
+} network;
+
+
+typedef struct box {
+ float x, y, w, h;
+} box;
+
+typedef struct detection{
+ box bbox;
+ float *prob;
+ float objectness;
+} detection;
+
+
+
+static int sort_class;
+
+static void free_dets(std::forward_list<detection> &dets){
+ std::forward_list<detection>::iterator it;
+ for ( it = dets.begin(); it != dets.end(); ++it ){
+ free(it->prob);
+ }
+}
+
+float sigmoid(float x)
+{
+ return 1.f/(1.f + exp(-x));
+}
+
+static bool det_objectness_comparator(detection &pa, detection &pb)
+{
+ return pa.objectness < pb.objectness;
+}
+
+static void insert_topN_det(std::forward_list<detection> &dets, detection det)
+{
+ std::forward_list<detection>::iterator it;
+ std::forward_list<detection>::iterator last_it;
+ for ( it = dets.begin(); it != dets.end(); ++it ){
+ if(it->objectness > det.objectness)
+ break;
+ last_it = it;
+ }
+ if(it != dets.begin()){
+ dets.emplace_after(last_it, det);
+ free(dets.begin()->prob);
+ dets.pop_front();
+ }
+ else{
+ free(det.prob);
+ }
+}
+
+static std::forward_list<detection> get_network_boxes(network *net, int image_w, int image_h, float thresh, int *num)
+{
+ std::forward_list<detection> dets;
+ int i;
+ int num_classes = net->num_classes;
+ *num = 0;
+
+ for (i = 0; i < net->num_branch; ++i) {
+ int height = net->branchs[i].resolution;
+ int width = net->branchs[i].resolution;
+ int channel = net->branchs[i].num_box*(5+num_classes);
+
+ for (int h = 0; h < net->branchs[i].resolution; h++) {
+ for (int w = 0; w < net->branchs[i].resolution; w++) {
+ for (int anc = 0; anc < net->branchs[i].num_box; anc++) {
+
+ // objectness score
+ int bbox_obj_offset = h * width * channel + w * channel + anc * (num_classes + 5) + 4;
+ float objectness = sigmoid(((float)net->branchs[i].tf_output[bbox_obj_offset] - net->branchs[i].zero_point) * net->branchs[i].scale);
+
+ if(objectness > thresh){
+ detection det;
+ det.prob = (float*)calloc(num_classes, sizeof(float));
+ det.objectness = objectness;
+ //get bbox prediction data for each anchor, each feature point
+ int bbox_x_offset = bbox_obj_offset -4;
+ int bbox_y_offset = bbox_x_offset + 1;
+ int bbox_w_offset = bbox_x_offset + 2;
+ int bbox_h_offset = bbox_x_offset + 3;
+ int bbox_scores_offset = bbox_x_offset + 5;
+ //int bbox_scores_step = 1;
+ det.bbox.x = ((float)net->branchs[i].tf_output[bbox_x_offset] - net->branchs[i].zero_point) * net->branchs[i].scale;
+ det.bbox.y = ((float)net->branchs[i].tf_output[bbox_y_offset] - net->branchs[i].zero_point) * net->branchs[i].scale;
+ det.bbox.w = ((float)net->branchs[i].tf_output[bbox_w_offset] - net->branchs[i].zero_point) * net->branchs[i].scale;
+ det.bbox.h = ((float)net->branchs[i].tf_output[bbox_h_offset] - net->branchs[i].zero_point) * net->branchs[i].scale;
+
+
+ float bbox_x, bbox_y;
+
+ // Eliminate grid sensitivity trick involved in YOLOv4
+ bbox_x = sigmoid(det.bbox.x); //* net->branchs[i].scale_x_y - (net->branchs[i].scale_x_y - 1) / 2;
+ bbox_y = sigmoid(det.bbox.y); //* net->branchs[i].scale_x_y - (net->branchs[i].scale_x_y - 1) / 2;
+ det.bbox.x = (bbox_x + w) / width;
+ det.bbox.y = (bbox_y + h) / height;
+
+ det.bbox.w = exp(det.bbox.w) * net->branchs[i].anchor[anc*2] / net->input_w;
+ det.bbox.h = exp(det.bbox.h) * net->branchs[i].anchor[anc*2+1] / net->input_h;
+
+ for (int s = 0; s < num_classes; s++) {
+ det.prob[s] = sigmoid(((float)net->branchs[i].tf_output[bbox_scores_offset + s] - net->branchs[i].zero_point) * net->branchs[i].scale)*objectness;
+ det.prob[s] = (det.prob[s] > thresh) ? det.prob[s] : 0;
+ }
+
+ //correct_yolo_boxes
+ det.bbox.x *= image_w;
+ det.bbox.w *= image_w;
+ det.bbox.y *= image_h;
+ det.bbox.h *= image_h;
+
+ if (*num < net->topN || net->topN <=0){
+ dets.emplace_front(det);
+ *num += 1;
+ }
+ else if(*num == net->topN){
+ dets.sort(det_objectness_comparator);
+ insert_topN_det(dets,det);
+ *num += 1;
+ }else{
+ insert_topN_det(dets,det);
+ }
+ }
+ }
+ }
+ }
+ }
+ if(*num > net->topN)
+ *num -=1;
+ return dets;
+}
+
+// init part
+
+static branch create_brach(int resolution, int num_box, float *anchor, int8_t *tf_output, size_t size, float scale, int zero_point)
+{
+ branch b;
+ b.resolution = resolution;
+ b.num_box = num_box;
+ b.anchor = anchor;
+ b.tf_output = tf_output;
+ b.size = size;
+ b.scale = scale;
+ b.zero_point = zero_point;
+ return b;
+}
+
+static network creat_network(int input_w, int input_h, int num_classes, int num_branch, branch* branchs, int topN)
+{
+ network net;
+ net.input_w = input_w;
+ net.input_h = input_h;
+ net.num_classes = num_classes;
+ net.num_branch = num_branch;
+ net.branchs = branchs;
+ net.topN = topN;
+ return net;
+}
+
+// NMS part
+
+static float Calc1DOverlap(float x1_center, float width1, float x2_center, float width2)
+{
+ float left_1 = x1_center - width1/2;
+ float left_2 = x2_center - width2/2;
+ float leftest;
+ if (left_1 > left_2) {
+ leftest = left_1;
+ } else {
+ leftest = left_2;
+ }
+
+ float right_1 = x1_center + width1/2;
+ float right_2 = x2_center + width2/2;
+ float rightest;
+ if (right_1 < right_2) {
+ rightest = right_1;
+ } else {
+ rightest = right_2;
+ }
+
+ return rightest - leftest;
+}
+
+
+static float CalcBoxIntersect(box box1, box box2)
+{
+ float width = Calc1DOverlap(box1.x, box1.w, box2.x, box2.w);
+ if (width < 0) return 0;
+ float height = Calc1DOverlap(box1.y, box1.h, box2.y, box2.h);
+ if (height < 0) return 0;
+
+ float total_area = width*height;
+ return total_area;
+}
+
+
+static float CalcBoxUnion(box box1, box box2)
+{
+ float boxes_intersection = CalcBoxIntersect(box1, box2);
+ float boxes_union = box1.w*box1.h + box2.w*box2.h - boxes_intersection;
+ return boxes_union;
+}
+
+
+static float CalcBoxIOU(box box1, box box2)
+{
+ float boxes_intersection = CalcBoxIntersect(box1, box2);
+
+ if (boxes_intersection == 0) return 0;
+
+ float boxes_union = CalcBoxUnion(box1, box2);
+
+ if (boxes_union == 0) return 0;
+
+ return boxes_intersection / boxes_union;
+}
+
+
+static bool CompareProbs(detection &prob1, detection &prob2)
+{
+ return prob1.prob[sort_class] > prob2.prob[sort_class];
+}
+
+
+static void CalcNMS(std::forward_list<detection> &detections, int classes, float iou_threshold)
+{
+ int k;
+
+ for (k = 0; k < classes; ++k) {
+ sort_class = k;
+ detections.sort(CompareProbs);
+
+ for (std::forward_list<detection>::iterator it=detections.begin(); it != detections.end(); ++it){
+ if (it->prob[k] == 0) continue;
+ for (std::forward_list<detection>::iterator itc=std::next(it, 1); itc != detections.end(); ++itc){
+ if (itc->prob[k] == 0) continue;
+ if (CalcBoxIOU(it->bbox, itc->bbox) > iou_threshold) {
+ itc->prob[k] = 0;
+ }
+ }
+ }
+ }
+}
+
+
+static void inline check_and_fix_offset(int im_w,int im_h,int *offset)
+{
+
+ if (!offset) return;
+
+ if ( (*offset) >= im_w*im_h*FORMAT_MULTIPLY_FACTOR)
+ (*offset) = im_w*im_h*FORMAT_MULTIPLY_FACTOR -1;
+ else if ( (*offset) < 0)
+ *offset =0;
+
+}
+
+
+static void DrawBoxOnImage(uint8_t *img_in,int im_w,int im_h,int bx,int by,int bw,int bh)
+{
+
+ if (!img_in) {
+ return;
+ }
+
+ int offset=0;
+ for (int i=0; i < bw; i++) {
+ /*draw two lines */
+ for (int line=0; line < 2; line++) {
+ /*top*/
+ offset =(i + (by + line)*im_w + bx)*FORMAT_MULTIPLY_FACTOR;
+ check_and_fix_offset(im_w,im_h,&offset);
+ img_in[offset] = 0xFF; /* FORMAT_MULTIPLY_FACTOR for rgb or grayscale*/
+ /*bottom*/
+ offset = (i + (by + bh - line)*im_w + bx)*FORMAT_MULTIPLY_FACTOR;
+ check_and_fix_offset(im_w,im_h,&offset);
+ img_in[offset] = 0xFF;
+ }
+ }
+
+ for (int i=0; i < bh; i++) {
+ /*draw two lines */
+ for (int line=0; line < 2; line++) {
+ /*left*/
+ offset = ((i + by)*im_w + bx + line)*FORMAT_MULTIPLY_FACTOR;
+ check_and_fix_offset(im_w,im_h,&offset);
+ img_in[offset] = 0xFF;
+ /*right*/
+ offset = ((i + by)*im_w + bx + bw - line)*FORMAT_MULTIPLY_FACTOR;
+ check_and_fix_offset(im_w,im_h,&offset);
+ img_in[offset] = 0xFF;
+ }
+ }
+
+}
+
+
+void arm::app::RunPostProcessing(uint8_t *img_in,TfLiteTensor* model_output[2],std::vector<arm::app::DetectionResult> & results_out)
+{
+
+ TfLiteTensor* output[2] = {nullptr,nullptr};
+ int input_w = INPUT_IMAGE_WIDTH;
+ int input_h = INPUT_IMAGE_HEIGHT;
+
+ for(int anchor=0;anchor<2;anchor++)
+ {
+ output[anchor] = model_output[anchor];
+ }
+
+ /* init postprocessing */
+ int num_classes = 1;
+ int num_branch = 2;
+ int topN = 0;
+
+ branch* branchs = (branch*)calloc(num_branch, sizeof(branch));
+
+ /*NOTE: anchors are different for any given input model size, estimated during training phase */
+ float anchor1[] = {38, 77, 47, 97, 61, 126};
+ float anchor2[] = {14, 26, 19, 37, 28, 55 };
+
+
+ branchs[0] = create_brach(INPUT_IMAGE_WIDTH/32, 3, anchor1, output[0]->data.int8, output[0]->bytes, ((TfLiteAffineQuantization*)(output[0]->quantization.params))->scale->data[0], ((TfLiteAffineQuantization*)(output[0]->quantization.params))->zero_point->data[0]);
+
+ branchs[1] = create_brach(INPUT_IMAGE_WIDTH/16, 3, anchor2, output[1]->data.int8, output[1]->bytes, ((TfLiteAffineQuantization*)(output[1]->quantization.params))->scale->data[0],((TfLiteAffineQuantization*)(output[1]->quantization.params))->zero_point->data[0]);
+
+ network net = creat_network(input_w, input_h, num_classes, num_branch, branchs,topN);
+ /* end init */
+
+ /* start postprocessing */
+ int nboxes=0;
+ float thresh = .5;//50%
+ float nms = .45;
+ int orig_image_width = ORIGINAL_IMAGE_WIDTH;
+ int orig_image_height = ORIGINAL_IMAGE_HEIGHT;
+ std::forward_list<detection> dets = get_network_boxes(&net, orig_image_width, orig_image_height, thresh, &nboxes);
+ /* do nms */
+ CalcNMS(dets, net.num_classes, nms);
+ uint8_t temp_unsuppressed_counter = 0;
+ int j;
+ for (std::forward_list<detection>::iterator it=dets.begin(); it != dets.end(); ++it){
+ float xmin = it->bbox.x - it->bbox.w / 2.0f;
+ float xmax = it->bbox.x + it->bbox.w / 2.0f;
+ float ymin = it->bbox.y - it->bbox.h / 2.0f;
+ float ymax = it->bbox.y + it->bbox.h / 2.0f;
+
+ if (xmin < 0) xmin = 0;
+ if (ymin < 0) ymin = 0;
+ if (xmax > orig_image_width) xmax = orig_image_width;
+ if (ymax > orig_image_height) ymax = orig_image_height;
+
+ float bx = xmin;
+ float by = ymin;
+ float bw = xmax - xmin;
+ float bh = ymax - ymin;
+
+ for (j = 0; j < net.num_classes; ++j) {
+ if (it->prob[j] > 0) {
+
+ arm::app::DetectionResult tmp_result = {};
+
+ tmp_result.m_normalisedVal = it->prob[j];
+ tmp_result.m_x0=bx;
+ tmp_result.m_y0=by;
+ tmp_result.m_w=bw;
+ tmp_result.m_h=bh;
+
+ results_out.push_back(tmp_result);
+
+ DrawBoxOnImage(img_in,orig_image_width,orig_image_height,bx,by,bw,bh);
+
+ temp_unsuppressed_counter++;
+ }
+ }
+ }
+
+ free_dets(dets);
+ free(branchs);
+
+}
+
+void arm::app::RgbToGrayscale(const uint8_t *rgb,uint8_t *gray, int im_w,int im_h)
+{
+ float R=0.299;
+ float G=0.587;
+ float B=0.114;
+ for (int i=0; i< im_w*im_h; i++ ) {
+
+ uint32_t int_gray = rgb[i*3 + 0]*R + rgb[i*3 + 1]*G+ rgb[i*3 + 2]*B;
+ /*clip if need */
+ if (int_gray <= UINT8_MAX) {
+ gray[i] = int_gray;
+ } else {
+ gray[i] = UINT8_MAX;
+ }
+
+ }
+
+}
+
diff --git a/source/use_case/object_detection/src/MainLoop.cc b/source/use_case/object_detection/src/MainLoop.cc
new file mode 100644
index 0000000..b0fbf96
--- /dev/null
+++ b/source/use_case/object_detection/src/MainLoop.cc
@@ -0,0 +1,83 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "hal.h" /* Brings in platform definitions. */
+#include "InputFiles.hpp" /* For input images. */
+#include "YoloFastestModel.hpp" /* Model class for running inference. */
+#include "UseCaseHandler.hpp" /* Handlers for different user options. */
+#include "UseCaseCommonUtils.hpp" /* Utils functions. */
+#include "DetectionUseCaseUtils.hpp" /* Utils functions specific to object detection. */
+
+
+void main_loop(hal_platform& platform)
+{
+ arm::app::YoloFastestModel model; /* Model wrapper object. */
+
+ /* Load the model. */
+ if (!model.Init()) {
+ printf_err("Failed to initialise model\n");
+ return;
+ }
+
+ /* Instantiate application context. */
+ arm::app::ApplicationContext caseContext;
+
+ arm::app::Profiler profiler{&platform, "object_detection"};
+ caseContext.Set<arm::app::Profiler&>("profiler", profiler);
+ caseContext.Set<hal_platform&>("platform", platform);
+ caseContext.Set<arm::app::Model&>("model", model);
+ caseContext.Set<uint32_t>("imgIndex", 0);
+
+
+ /* Loop. */
+ bool executionSuccessful = true;
+ constexpr bool bUseMenu = NUMBER_OF_FILES > 1 ? true : false;
+
+ /* Loop. */
+ do {
+ int menuOption = common::MENU_OPT_RUN_INF_NEXT;
+ if (bUseMenu) {
+ DisplayDetectionMenu();
+ menuOption = arm::app::ReadUserInputAsInt(platform);
+ printf("\n");
+ }
+ switch (menuOption) {
+ case common::MENU_OPT_RUN_INF_NEXT:
+ executionSuccessful = ObjectDetectionHandler(caseContext, caseContext.Get<uint32_t>("imgIndex"), false);
+ break;
+ case common::MENU_OPT_RUN_INF_CHOSEN: {
+ printf(" Enter the image index [0, %d]: ", NUMBER_OF_FILES-1);
+ fflush(stdout);
+ auto imgIndex = static_cast<uint32_t>(arm::app::ReadUserInputAsInt(platform));
+ executionSuccessful = ObjectDetectionHandler(caseContext, imgIndex, false);
+ break;
+ }
+ case common::MENU_OPT_RUN_INF_ALL:
+ executionSuccessful = ObjectDetectionHandler(caseContext, caseContext.Get<uint32_t>("imgIndex"), true);
+ break;
+ case common::MENU_OPT_SHOW_MODEL_INFO:
+ executionSuccessful = model.ShowModelInfoHandler();
+ break;
+ case common::MENU_OPT_LIST_IFM:
+ executionSuccessful = ListFilesHandler(caseContext);
+ break;
+ default:
+ printf("Incorrect choice, try again.");
+ break;
+ }
+ } while (executionSuccessful && bUseMenu);
+ info("Main loop terminated.\n");
+}
diff --git a/source/use_case/object_detection/src/UseCaseHandler.cc b/source/use_case/object_detection/src/UseCaseHandler.cc
new file mode 100644
index 0000000..45df4f8
--- /dev/null
+++ b/source/use_case/object_detection/src/UseCaseHandler.cc
@@ -0,0 +1,162 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "UseCaseHandler.hpp"
+#include "InputFiles.hpp"
+#include "YoloFastestModel.hpp"
+#include "UseCaseCommonUtils.hpp"
+#include "DetectionUseCaseUtils.hpp"
+#include "DetectorPostProcessing.hpp"
+#include "hal.h"
+
+#include <inttypes.h>
+
+
+/* used for presentation, original images are read-only"*/
+static uint8_t g_image_buffer[INPUT_IMAGE_WIDTH*INPUT_IMAGE_HEIGHT*FORMAT_MULTIPLY_FACTOR] IFM_BUF_ATTRIBUTE = {};
+
+namespace arm {
+namespace app {
+
+
+ /* Object detection classification handler. */
+ bool ObjectDetectionHandler(ApplicationContext& ctx, uint32_t imgIndex, bool runAll)
+ {
+ auto& platform = ctx.Get<hal_platform&>("platform");
+ auto& profiler = ctx.Get<Profiler&>("profiler");
+
+ constexpr uint32_t dataPsnImgDownscaleFactor = 1;
+ constexpr uint32_t dataPsnImgStartX = 10;
+ constexpr uint32_t dataPsnImgStartY = 35;
+
+ constexpr uint32_t dataPsnTxtInfStartX = 150;
+ constexpr uint32_t dataPsnTxtInfStartY = 40;
+
+ platform.data_psn->clear(COLOR_BLACK);
+
+ auto& model = ctx.Get<Model&>("model");
+
+ /* If the request has a valid size, set the image index. */
+ if (imgIndex < NUMBER_OF_FILES) {
+ if (!SetAppCtxIfmIdx(ctx, imgIndex, "imgIndex")) {
+ return false;
+ }
+ }
+ if (!model.IsInited()) {
+ printf_err("Model is not initialised! Terminating processing.\n");
+ return false;
+ }
+
+ auto curImIdx = ctx.Get<uint32_t>("imgIndex");
+
+ TfLiteTensor* inputTensor = model.GetInputTensor(0);
+
+ if (!inputTensor->dims) {
+ printf_err("Invalid input tensor dims\n");
+ return false;
+ } else if (inputTensor->dims->size < 3) {
+ printf_err("Input tensor dimension should be >= 3\n");
+ return false;
+ }
+
+ TfLiteIntArray* inputShape = model.GetInputShape(0);
+
+ const uint32_t nCols = inputShape->data[arm::app::YoloFastestModel::ms_inputColsIdx];
+ const uint32_t nRows = inputShape->data[arm::app::YoloFastestModel::ms_inputRowsIdx];
+ const uint32_t nPresentationChannels = FORMAT_MULTIPLY_FACTOR;
+
+ std::vector<DetectionResult> results;
+
+ do {
+ /* Strings for presentation/logging. */
+ std::string str_inf{"Running inference... "};
+
+ const uint8_t* curr_image = get_img_array(ctx.Get<uint32_t>("imgIndex"));
+
+ /* Copy over the data and convert to gryscale */
+#if DISPLAY_RGB_IMAGE
+ memcpy(g_image_buffer,curr_image, INPUT_IMAGE_WIDTH*INPUT_IMAGE_HEIGHT*FORMAT_MULTIPLY_FACTOR);
+#else
+ RgbToGrayscale(curr_image,g_image_buffer,INPUT_IMAGE_WIDTH,INPUT_IMAGE_HEIGHT);
+#endif /*DISPLAY_RGB_IMAGE*/
+
+ RgbToGrayscale(curr_image,inputTensor->data.uint8,INPUT_IMAGE_WIDTH,INPUT_IMAGE_HEIGHT);
+
+
+ /* Display this image on the LCD. */
+ platform.data_psn->present_data_image(
+ g_image_buffer,
+ nCols, nRows, nPresentationChannels,
+ dataPsnImgStartX, dataPsnImgStartY, dataPsnImgDownscaleFactor);
+
+ /* If the data is signed. */
+ if (model.IsDataSigned()) {
+ image::ConvertImgToInt8(inputTensor->data.data, inputTensor->bytes);
+ }
+
+ /* Display message on the LCD - inference running. */
+ platform.data_psn->present_data_text(str_inf.c_str(), str_inf.size(),
+ dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0);
+
+ /* Run inference over this image. */
+ info("Running inference on image %" PRIu32 " => %s\n", ctx.Get<uint32_t>("imgIndex"),
+ get_filename(ctx.Get<uint32_t>("imgIndex")));
+
+ if (!RunInference(model, profiler)) {
+ return false;
+ }
+
+ /* Erase. */
+ str_inf = std::string(str_inf.size(), ' ');
+ platform.data_psn->present_data_text(str_inf.c_str(), str_inf.size(),
+ dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0);
+
+ /* Detector post-processing*/
+ TfLiteTensor* output_arr[2] = {nullptr,nullptr};
+ output_arr[0] = model.GetOutputTensor(0);
+ output_arr[1] = model.GetOutputTensor(1);
+ RunPostProcessing(g_image_buffer,output_arr,results);
+
+ platform.data_psn->present_data_image(
+ g_image_buffer,
+ nCols, nRows, nPresentationChannels,
+ dataPsnImgStartX, dataPsnImgStartY, dataPsnImgDownscaleFactor);
+
+ /*Detector post-processing*/
+
+
+ /* Add results to context for access outside handler. */
+ ctx.Set<std::vector<DetectionResult>>("results", results);
+
+#if VERIFY_TEST_OUTPUT
+ arm::app::DumpTensor(outputTensor);
+#endif /* VERIFY_TEST_OUTPUT */
+
+ if (!image::PresentInferenceResult(platform, results)) {
+ return false;
+ }
+
+ profiler.PrintProfilingResult();
+
+ IncrementAppCtxIfmIdx(ctx,"imgIndex");
+
+ } while (runAll && ctx.Get<uint32_t>("imgIndex") != curImIdx);
+
+ return true;
+ }
+
+} /* namespace app */
+} /* namespace arm */
diff --git a/source/use_case/object_detection/src/YoloFastestModel.cc b/source/use_case/object_detection/src/YoloFastestModel.cc
new file mode 100644
index 0000000..a8afd59
--- /dev/null
+++ b/source/use_case/object_detection/src/YoloFastestModel.cc
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "YoloFastestModel.hpp"
+
+#include "hal.h"
+
+const tflite::MicroOpResolver& arm::app::YoloFastestModel::GetOpResolver()
+{
+ return this->m_opResolver;
+}
+
+bool arm::app::YoloFastestModel::EnlistOperations()
+{
+ this->m_opResolver.AddDepthwiseConv2D();
+ this->m_opResolver.AddConv2D();
+ this->m_opResolver.AddAdd();
+ this->m_opResolver.AddResizeNearestNeighbor();
+ /*These are needed for UT to work, not needed on FVP */
+ this->m_opResolver.AddPad();
+ this->m_opResolver.AddMaxPool2D();
+ this->m_opResolver.AddConcatenation();
+
+#if defined(ARM_NPU)
+ if (kTfLiteOk == this->m_opResolver.AddEthosU()) {
+ info("Added %s support to op resolver\n",
+ tflite::GetString_ETHOSU());
+ } else {
+ printf_err("Failed to add Arm NPU support to op resolver.");
+ return false;
+ }
+#endif /* ARM_NPU */
+ return true;
+}
+
+extern uint8_t* GetModelPointer();
+const uint8_t* arm::app::YoloFastestModel::ModelPointer()
+{
+ return GetModelPointer();
+}
+
+extern size_t GetModelLen();
+size_t arm::app::YoloFastestModel::ModelSize()
+{
+ return GetModelLen();
+}
diff --git a/source/use_case/object_detection/usecase.cmake b/source/use_case/object_detection/usecase.cmake
new file mode 100644
index 0000000..15bf534
--- /dev/null
+++ b/source/use_case/object_detection/usecase.cmake
@@ -0,0 +1,59 @@
+#----------------------------------------------------------------------------
+# Copyright (c) 2022 Arm Limited. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#----------------------------------------------------------------------------
+
+USER_OPTION(${use_case}_FILE_PATH "Directory with custom image files to use, or path to a single image, in the evaluation application"
+ ${CMAKE_CURRENT_SOURCE_DIR}/resources/${use_case}/samples/
+ PATH_OR_FILE)
+
+USER_OPTION(${use_case}_IMAGE_SIZE "Square image size in pixels. Images will be resized to this size."
+ 192
+ STRING)
+
+add_compile_definitions(DISPLAY_RGB_IMAGE=1)
+add_compile_definitions(INPUT_IMAGE_WIDTH=${${use_case}_IMAGE_SIZE})
+add_compile_definitions(INPUT_IMAGE_HEIGHT=${${use_case}_IMAGE_SIZE})
+add_compile_definitions(ORIGINAL_IMAGE_WIDTH=${${use_case}_IMAGE_SIZE})
+add_compile_definitions(ORIGINAL_IMAGE_HEIGHT=${${use_case}_IMAGE_SIZE})
+
+
+# Generate input files
+generate_images_code("${${use_case}_FILE_PATH}"
+ ${SRC_GEN_DIR}
+ ${INC_GEN_DIR}
+ "${${use_case}_IMAGE_SIZE}")
+
+
+USER_OPTION(${use_case}_ACTIVATION_BUF_SZ "Activation buffer size for the chosen model"
+ 0x00082000
+ STRING)
+
+if (ETHOS_U_NPU_ENABLED)
+ set(DEFAULT_MODEL_PATH ${DEFAULT_MODEL_DIR}/yolo-fastest_192_face_v4_vela_${ETHOS_U_NPU_CONFIG_ID}.tflite)
+else()
+ set(DEFAULT_MODEL_PATH ${DEFAULT_MODEL_DIR}/yolo-fastest_192_face_v4.tflite)
+endif()
+
+USER_OPTION(${use_case}_MODEL_TFLITE_PATH "NN models file to be used in the evaluation application. Model files must be in tflite format."
+ ${DEFAULT_MODEL_PATH}
+ FILEPATH
+ )
+
+# Generate model file
+generate_tflite_code(
+ MODEL_PATH ${${use_case}_MODEL_TFLITE_PATH}
+ DESTINATION ${SRC_GEN_DIR}
+ )