From ed35a6fea4a1604db81c56fc71f7756822fcf212 Mon Sep 17 00:00:00 2001 From: Richard Burton Date: Mon, 14 Feb 2022 11:55:35 +0000 Subject: MLECO-2874: Move NMS out of the OD use_case * Add ImageUtils * Move image related code from UseCaseCommonUtils to ImageUtils * Move NMS related code to ImageUtils * Delete test specific ImageUtils and use new ImageUtils Signed-off-by: Richard Burton Change-Id: Icbf5dd9c6a941b0126ecdf69a0c9d9969f22729f --- source/application/main/ImageUtils.cc | 126 +++++++++++++++++++++ source/application/main/UseCaseCommonUtils.cc | 25 +--- source/application/main/include/ImageUtils.hpp | 116 +++++++++++++++++++ .../main/include/UseCaseCommonUtils.hpp | 32 +----- 4 files changed, 246 insertions(+), 53 deletions(-) create mode 100644 source/application/main/ImageUtils.cc create mode 100644 source/application/main/include/ImageUtils.hpp (limited to 'source/application') diff --git a/source/application/main/ImageUtils.cc b/source/application/main/ImageUtils.cc new file mode 100644 index 0000000..31b9493 --- /dev/null +++ b/source/application/main/ImageUtils.cc @@ -0,0 +1,126 @@ +/* + * 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 "ImageUtils.hpp" + +#include + +namespace arm { +namespace app { +namespace image { + + float Calculate1DOverlap(float x1Center, float width1, float x2Center, float width2) + { + float left_1 = x1Center - width1/2; + float left_2 = x2Center - width2/2; + float leftest = left_1 > left_2 ? left_1 : left_2; + + float right_1 = x1Center + width1/2; + float right_2 = x2Center + width2/2; + float rightest = right_1 < right_2 ? right_1 : right_2; + + return rightest - leftest; + } + + float CalculateBoxIntersect(Box& box1, Box& box2) + { + float width = Calculate1DOverlap(box1.x, box1.w, box2.x, box2.w); + if (width < 0) { + return 0; + } + float height = Calculate1DOverlap(box1.y, box1.h, box2.y, box2.h); + if (height < 0) { + return 0; + } + + float total_area = width*height; + return total_area; + } + + float CalculateBoxUnion(Box& box1, Box& box2) + { + float boxes_intersection = CalculateBoxIntersect(box1, box2); + float boxes_union = box1.w * box1.h + box2.w * box2.h - boxes_intersection; + return boxes_union; + } + + float CalculateBoxIOU(Box& box1, Box& box2) + { + float boxes_intersection = CalculateBoxIntersect(box1, box2); + if (boxes_intersection == 0) { + return 0; + } + + float boxes_union = CalculateBoxUnion(box1, box2); + if (boxes_union == 0) { + return 0; + } + + return boxes_intersection / boxes_union; + } + + void CalculateNMS(std::forward_list& detections, int classes, float iouThreshold) + { + int idxClass{0}; + auto CompareProbs = [idxClass](Detection& prob1, Detection& prob2) { + return prob1.prob[idxClass] > prob2.prob[idxClass]; + }; + + for (idxClass = 0; idxClass < classes; ++idxClass) { + detections.sort(CompareProbs); + + for (auto it=detections.begin(); it != detections.end(); ++it) { + if (it->prob[idxClass] == 0) continue; + for (auto itc=std::next(it, 1); itc != detections.end(); ++itc) { + if (itc->prob[idxClass] == 0) { + continue; + } + if (CalculateBoxIOU(it->bbox, itc->bbox) > iouThreshold) { + itc->prob[idxClass] = 0; + } + } + } + } + } + + void ConvertImgToInt8(void* data, const size_t kMaxImageSize) + { + auto* tmp_req_data = static_cast(data); + auto* tmp_signed_req_data = static_cast(data); + + for (size_t i = 0; i < kMaxImageSize; i++) { + tmp_signed_req_data[i] = (int8_t) ( + (int32_t) (tmp_req_data[i]) - 128); + } + } + + void RgbToGrayscale(const uint8_t* srcPtr, uint8_t* dstPtr, const size_t dstImgSz) + { + const float R = 0.299; + const float G = 0.587; + const float B = 0.114; + for (size_t i = 0; i < dstImgSz; ++i, srcPtr += 3) { + uint32_t int_gray = R * (*srcPtr) + + G * (*(srcPtr + 1)) + + B * (*(srcPtr + 2)); + *dstPtr++ = int_gray <= std::numeric_limits::max() ? + int_gray : std::numeric_limits::max(); + } + } + +} /* namespace image */ +} /* namespace app */ +} /* namespace arm */ \ No newline at end of file diff --git a/source/application/main/UseCaseCommonUtils.cc b/source/application/main/UseCaseCommonUtils.cc index dd9a32d..d439446 100644 --- a/source/application/main/UseCaseCommonUtils.cc +++ b/source/application/main/UseCaseCommonUtils.cc @@ -15,6 +15,7 @@ * limitations under the License. */ #include "UseCaseCommonUtils.hpp" +#include "ImageUtils.hpp" #include "InputFiles.hpp" #include "log_macros.h" @@ -35,19 +36,10 @@ void DisplayCommonMenu() fflush(stdout); } -void image::ConvertImgToInt8(void* data, const size_t kMaxImageSize) -{ - auto* tmp_req_data = static_cast(data); - auto* tmp_signed_req_data = static_cast(data); - for (size_t i = 0; i < kMaxImageSize; i++) { - tmp_signed_req_data[i] = (int8_t) ( - (int32_t) (tmp_req_data[i]) - 128); - } -} -bool image::PresentInferenceResult( +bool PresentInferenceResult( hal_platform &platform, const std::vector &results) { @@ -93,19 +85,6 @@ bool image::PresentInferenceResult( return true; } -void image::RgbToGrayscale(const uint8_t *srcPtr, uint8_t *dstPtr, const size_t dstImgSz) -{ - float R=0.299; - float G=0.587; - float B=0.114; - for (size_t i = 0; i < dstImgSz; ++i, srcPtr += 3) { - uint32_t int_gray = R * (*srcPtr) + - G * (*(srcPtr + 1)) + - B * (*(srcPtr + 2)); - *dstPtr++ = int_gray <= std::numeric_limits::max() ? - int_gray : std::numeric_limits::max(); - } -} void IncrementAppCtxIfmIdx(arm::app::ApplicationContext& ctx, const std::string& useCase) { diff --git a/source/application/main/include/ImageUtils.hpp b/source/application/main/include/ImageUtils.hpp new file mode 100644 index 0000000..a8c7650 --- /dev/null +++ b/source/application/main/include/ImageUtils.hpp @@ -0,0 +1,116 @@ +/* + * 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 IMAGE_UTILS_HPP +#define IMAGE_UTILS_HPP + +#include +#include +#include +#include + +/* Helper macro to convert RGB888 to RGB565 format. */ +#define RGB888_TO_RGB565(R8,G8,B8) ((((R8>>3) & 0x1F) << 11) | \ + (((G8>>2) & 0x3F) << 5) | \ + ((B8>>3) & 0x1F)) + +constexpr uint16_t COLOR_BLACK = 0; +constexpr uint16_t COLOR_GREEN = RGB888_TO_RGB565( 0, 255, 0); // 2016; +constexpr uint16_t COLOR_YELLOW = RGB888_TO_RGB565(255, 255, 0); // 65504; + + +namespace arm { +namespace app { +namespace image { + + /** + * Contains the x,y co-ordinates of a box centre along with the box width and height. + */ + struct Box { + float x; + float y; + float w; + float h; + }; + + struct Detection { + Box bbox; + std::vector prob; + float objectness; + }; + + /** + * @brief Calculate the 1D overlap. + * @param[in] x1Center First center point. + * @param[in] width1 First width. + * @param[in] x2Center Second center point. + * @param[in] width2 Second width. + * @return The overlap between the two lines. + **/ + float Calculate1DOverlap(float x1Center, float width1, float x2Center, float width2); + + /** + * @brief Calculate the intersection between the two given boxes. + * @param[in] box1 First box. + * @param[in] box2 Second box. + * @return The intersection value. + **/ + float CalculateBoxIntersect(Box& box1, Box& box2); + + /** + * @brief Calculate the union between the two given boxes. + * @param[in] box1 First box. + * @param[in] box2 Second box. + * @return The two given boxes union value. + **/ + float CalculateBoxUnion(Box& box1, Box& box2); + + /** + * @brief Calculate the intersection over union between the two given boxes. + * @param[in] box1 First box. + * @param[in] box2 Second box. + * @return The intersection over union value. + **/ + float CalculateBoxIOU(Box& box1, Box& box2); + + /** + * @brief Calculate the Non-Maxima suppression on the given detection boxes. + * @param[in] detections List of Detection boxes. + * @param[in] classes Number of classes. + * @param[in] iouThreshold Intersection over union threshold. + **/ + void CalculateNMS(std::forward_list& detections, int classes, float iouThreshold); + + /** + * @brief Helper function to convert a UINT8 image to INT8 format. + * @param[in,out] data Pointer to the data start. + * @param[in] kMaxImageSize Total number of pixels in the image. + **/ + void ConvertImgToInt8(void* data, size_t kMaxImageSize); + + /** + * @brief Converts RGB image to grayscale. + * @param[in] srcPtr Pointer to RGB source image. + * @param[out] dstPtr Pointer to grayscale destination image. + * @param[in] imgSz Destination image size. + **/ + void RgbToGrayscale(const uint8_t* srcPtr, uint8_t* dstPtr, size_t dstImgSz); + +} /* namespace image */ +} /* namespace app */ +} /* namespace arm */ + +#endif /* IMAGE_UTILS_HPP */ \ No newline at end of file diff --git a/source/application/main/include/UseCaseCommonUtils.hpp b/source/application/main/include/UseCaseCommonUtils.hpp index cd0cb69..7f5dde6 100644 --- a/source/application/main/include/UseCaseCommonUtils.hpp +++ b/source/application/main/include/UseCaseCommonUtils.hpp @@ -24,30 +24,10 @@ #include "UseCaseHandler.hpp" /* Handlers for different user options. */ #include "Classifier.hpp" /* Classifier. */ #include "InputFiles.hpp" -#include - - -/* Helper macro to convert RGB888 to RGB565 format. */ -#define RGB888_TO_RGB565(R8,G8,B8) ((((R8>>3) & 0x1F) << 11) | \ - (((G8>>2) & 0x3F) << 5) | \ - ((B8>>3) & 0x1F)) - -constexpr uint16_t COLOR_BLACK = 0; -constexpr uint16_t COLOR_GREEN = RGB888_TO_RGB565( 0u, 255u, 0u); // 2016; -constexpr uint16_t COLOR_YELLOW = RGB888_TO_RGB565(255u, 255u, 0u); // 65504; void DisplayCommonMenu(); -namespace image{ - - /** - * @brief Helper function to convert a UINT8 image to INT8 format. - * @param[in,out] data Pointer to the data start. - * @param[in] kMaxImageSize Total number of pixels in the image. - **/ - void ConvertImgToInt8(void * data, size_t kMaxImageSize); - /** * @brief Presents inference results using the data presentation * object. @@ -55,17 +35,9 @@ namespace image{ * @param[in] results Vector of classification results to be displayed. * @return true if successful, false otherwise. **/ - bool PresentInferenceResult(hal_platform & platform, - const std::vector < arm::app::ClassificationResult > & results); +bool PresentInferenceResult(hal_platform& platform, + const std::vector& results); - /** - * @brief Converts RGB image to grayscale. - * @param[in] srcPtr Pointer to RGB source image. - * @param[out] dstPtr Pointer to grayscale destination image. - * @param[in] imgSz Destination image size. - **/ - void RgbToGrayscale(const uint8_t *srcPtr, uint8_t *dstPtr, const size_t dstImgSz); -} /** * @brief Helper function to increment current input feature vector index. -- cgit v1.2.1