misc

This reverts commit e5cf60e2547db7e7b27d03080b16bf1db79807bf.

CMakeLists.txt

@@ -1,17 +1,58 @@
 cmake_minimum_required(VERSION 3.12)
-
 project(match)
+
 find_package(OpenCV REQUIRED)
-add_executable(${PROJECT_NAME}
-    main.cpp
+
+option(ENABLE_OPENMP "enable openmp" OFF)
+if(ENABLE_OPENMP)
+        # find OpenMP
+        find_package(OpenMP REQUIRED)
+
+        if(OPENMP_FOUND)
+            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
+            set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
+        endif(OPENMP_FOUND)
+endif(ENABLE_OPENMP)
+
+#==============================================================
+#library
+#==============================================================
+add_library(algo SHARED
     grayMatch.h
     grayMatch.cpp
+    sum.h
+    sum.cpp
+    privateType.h
+    apiExport.h
 )
 
-target_include_directories(${PROJECT_NAME} PRIVATE ${OpenCV_INCLUDE_DIRS})
-target_link_libraries(${PROJECT_NAME} ${OpenCV_LIBRARIES})
-target_compile_options(${PROJECT_NAME} PRIVATE
+target_include_directories(algo PRIVATE ${OpenCV_INCLUDE_DIRS})
+target_link_libraries(algo ${OpenCV_LIBRARIES} $<$<BOOL:${OPENMP_FOUND}>:OpenMP::OpenMP_CXX>)
+target_compile_options(algo PRIVATE
         $<$<CXX_COMPILER_ID:MSVC>:/W4 /WX /external:W0>
         $<$<STREQUAL:${CMAKE_SYSTEM_NAME},Linux>:-fPIC -fvisibility=hidden -Wall -Wextra -Wpedantic -Wmisleading-indentation -Wunused -Wuninitialized -Wshadow -Wconversion -Werror>
         $<$<AND:$<CXX_COMPILER_ID:Clang>,$<STREQUAL:${CMAKE_SYSTEM_NAME},Windows>>:/W4 /WX /external:W0>
+        $<$<STREQUAL:${CMAKE_SYSTEM_PROCESSOR},loongarch64>:-mlsx>
 )
+target_compile_definitions(algo PRIVATE  API_EXPORTS
+        $<$<STREQUAL:${CMAKE_SYSTEM_PROCESSOR},loongarch64>:CV_LSX>
+)
+
+#==============================================================
+#exe
+#==============================================================
+add_executable(${PROJECT_NAME}
+    main.cpp
+)
+
+target_include_directories(${PROJECT_NAME} PRIVATE ${OpenCV_INCLUDE_DIRS})
+target_link_libraries(${PROJECT_NAME} ${OpenCV_LIBRARIES} algo)
+target_compile_options(${PROJECT_NAME} PRIVATE
+        $<$<CXX_COMPILER_ID:MSVC>:/W4 /WX /external:W0>
+        $<$<STREQUAL:${CMAKE_SYSTEM_NAME},Linux>: -fPIC -fvisibility=hidden -Wall -Wextra -Wpedantic -Wmisleading-indentation -Wunused -Wuninitialized -Wshadow -Wconversion -Werror>
+        $<$<AND:$<CXX_COMPILER_ID:Clang>,$<STREQUAL:${CMAKE_SYSTEM_NAME},Windows>>:/W4 /WX /external:W0>
+)
+target_compile_definitions(${PROJECT_NAME} PRIVATE IMG_DIR="${CMAKE_CURRENT_SOURCE_DIR}/img")
+
+message("Arch:${CMAKE_SYSTEM_PROCESSOR}")

README.md

@@ -0,0 +1,13 @@
+# Template match with gray model(ncc)
+
+## rotate-model vs main branch
+
+
+| method    | main(ms) | rotate-model(ms) | factor(main/rotate-model) |
+|-----------|----------|------------------|---------------------------|
+| train     | 1        | 680              | 1/680                     |     
+| match     | 31       | 16               | 2                         |
+| train-omp | 1        | 160              | 1/160                     |
+| match-omp | 12       | 6                | 2                         |
+
+result: main brain method train model really fast(1ms) and parameter free, rotate-model method 2x faster in matching.

apiExport.h

@@ -0,0 +1,31 @@
+#pragma once
+
+#if defined(_WIN32) || defined(_WIN64) || defined(__WINDOWS__)
+#define API_EXPORT __declspec(dllexport)
+#define API_IMPORT __declspec(dllimport)
+#define API_LOCAL
+#elif defined(linux) || defined(__linux) || defined(__linux__)
+#define API_EXPORT __attribute__((visibility("default")))
+#define API_IMPORT __attribute__((visibility("default")))
+#define API_LOCAL __attribute__((visibility("hidden")))
+#elif defined(__APPLE__)
+#define API_EXPORT __attribute__((visibility("default")))
+#define API_IMPORT __attribute__((visibility("default")))
+#define API_LOCAL __attribute__((visibility("hidden")))
+#else
+#define API_EXPORT
+#define API_IMPORT
+#define API_LOCAL
+#endif
+
+#ifdef __cplusplus
+#define API_DEMANGLED extern "C"
+#else
+#define API_DEMANGLED
+#endif
+
+#ifdef API_EXPORTS
+#define API_PUBLIC API_DEMANGLED API_EXPORT
+#else
+#define API_PUBLIC API_DEMANGLED API_IMPORT
+#endif

grayMatch.h

@@ -3,6 +3,8 @@
 
 #include <opencv2/opencv.hpp>
 
+#include "apiExport.h"
+
 struct Model;
 
 struct Pose {
@@ -12,14 +14,11 @@ struct Pose {
     float score;
 };
 
-Model *trainModel(const cv::Mat &src, int level);
+API_PUBLIC Model *trainModel(const cv::Mat &src, int level, double startAngle, double spanAngle,
+                             double angleStep);
 
-std::vector<Pose> matchModel(const cv::Mat &dst, const Model *model, int level, double startAngle,
-                             double spanAngle, double maxOverlap, double minScore, int maxCount,
-                             int subpixel);
-
-void serialize(Model *model, int &size, uint8_t *buffer);
-
-Model *deserialize(int size, uint8_t *buffer);
+API_PUBLIC void matchModel(const cv::Mat &dst, const Model *model, int *count, Pose *poses,
+                           int level, double startAngle, double spanAngle, double maxOverlap,
+                           double minScore, int subpixel);
 
 #endif // GRAY_MATCH_H

main.cpp

@@ -1,42 +1,93 @@
 #include "grayMatch.h"
 
 #include <iostream>
+#include <opencv2/core/utility.hpp>
 #include <opencv2/opencv.hpp>
+#include <string>
 
-int main() {
-    auto src =
-        cv::imread("C:/Users/qiuyong/Desktop/test/template/model3.bmp", cv::IMREAD_GRAYSCALE);
-    auto dst =
-        cv::imread("C:/Users/qiuyong/Desktop/test/template/model3_src1.bmp", cv::IMREAD_GRAYSCALE);
+int main(int argc, const char *argv[]) {
+    const std::string keys = "{model m || model image}"
+                             "{scene s || scene image}"
+                             "{view v || view result}"
+                             "{threshold t | 0.7 | match minium score}"
+                             "{bench b || match benchmark}"
+                             "{help h || print this help}";
+
+    cv::CommandLineParser cmd(argc, argv, keys);
+    if (!cmd.check()) {
+        cmd.printErrors();
+        return -1;
+    }
+
+    if (cmd.has("help")) {
+        cmd.printMessage();
+        return 0;
+    }
+
+    auto srcFile = std::string(IMG_DIR) + "/model3.png";
+    auto dstFile = std::string(IMG_DIR) + "/model3_src2.png";
+    if (cmd.has("model"))
+        srcFile = cmd.get<std::string>("model");
+    if (cmd.has("scene"))
+        dstFile = cmd.get<std::string>("scene");
+
+    auto src = cv::imread(srcFile, cv::IMREAD_GRAYSCALE);
+    auto dst = cv::imread(dstFile, cv::IMREAD_GRAYSCALE);
+    if (src.empty() || dst.empty()) {
+        return -1;
+    }
+
+    int               count = 70;
+    std::vector<Pose> poses(count);
+    auto              score = cmd.get<float>("threshold");
 
     auto t0    = cv::getTickCount();
-    auto model = trainModel(src, -1);
+    auto model = trainModel(src, -1, 0, 360, -1);
     auto t1    = cv::getTickCount();
-    auto poses = matchModel(dst, model, -1, 0, 360, 0, 0.5, 70, 1);
+    matchModel(dst, model, &count, poses.data(), -1, 0, 360, 0, score, 1);
     auto t2 = cv::getTickCount();
 
-    auto trainCost = double(t1 - t0) / cv::getTickFrequency();
-    auto matchCost = double(t2 - t1) / cv::getTickFrequency();
+    const auto trainCost = static_cast<double>(t1 - t0) / cv::getTickFrequency();
+    const auto matchCost = static_cast<double>(t2 - t1) / cv::getTickFrequency();
     std::cout << "train(s):" << trainCost << " match(s):" << matchCost << std::endl;
+    for (int i = 0; i < count; i++) {
+        const auto &pose = poses[ i ];
+        std::cout << pose.x << "," << pose.y << "," << pose.angle << "," << pose.score << std::endl;
+    }
 
+    if (cmd.has("bench")) {
+        constexpr int times = 100;
+
+        auto start = cv::getTickCount();
+        for (int i = 0; i < times; i++) {
+            matchModel(dst, model, &count, poses.data(), -1, 0, 360, 0, score, 1);
+            count = 70;
+        }
+        auto end = cv::getTickCount();
+
+        const auto cost = static_cast<double>(end - start) / cv::getTickFrequency() / times;
+        std::cout << "match bench avg(s):" << cost << std::endl;
+    }
+
+    if (cmd.has("view")) {
         cv::Mat color;
         cv::cvtColor(dst, color, cv::COLOR_GRAY2RGB);
-    for (auto &pose : poses) {
+        for (int i = 0; i < count; i++) {
+            const auto     &pose = poses[ i ];
             cv::RotatedRect rect(cv::Point2f(pose.x, pose.y), src.size(), -pose.angle);
 
-        std::vector<cv::Point2f> pts;
+            cv::Point2f pts[ 4 ];
             rect.points(pts);
 
             cv::line(color, pts[ 0 ], pts[ 1 ], cv::Scalar(255, 0, 0), 1, cv::LINE_AA);
             cv::line(color, pts[ 1 ], pts[ 2 ], cv::Scalar(255, 0, 0), 1, cv::LINE_AA);
             cv::line(color, pts[ 2 ], pts[ 3 ], cv::Scalar(255, 0, 0), 1, cv::LINE_AA);
             cv::line(color, pts[ 3 ], pts[ 0 ], cv::Scalar(255, 0, 0), 1, cv::LINE_AA);
-
-        std::cout << pose.x << "," << pose.y << "," << pose.angle << "," << pose.score << std::endl;
         }
 
         cv::imshow("img", color);
         cv::waitKey();
+    }
 
     return 0;
 }

privateType.h

@@ -0,0 +1,50 @@
+#pragma once
+
+#include <opencv2/core.hpp>
+
+struct HRLE {
+    int row         = -1;
+    int startColumn = -1;
+    int length      = 0;
+};
+
+struct VRLE {
+    int col      = -1;
+    int startRow = -1;
+    int length   = 0;
+};
+
+using HRegion = std::vector<HRLE>;
+using VRegion = std::vector<VRLE>;
+
+struct Template {
+    cv::Mat         img;
+    HRegion         hRegion;
+    VRegion         vRegion;
+    cv::RotatedRect rect;
+
+    double mean    = 0;
+    double normal  = 0;
+    double invArea = 0;
+};
+
+struct Layer {
+    double angleStep = 0;
+
+    std::vector<Template> templates;
+};
+
+struct Model {
+    double startAngle = 0;
+    double stopAngle  = 0;
+    double angleStep  = 0;
+
+    cv::Size           srcSize;
+    std::vector<Layer> layers;
+};
+
+#if CV_VERSION_MAJOR >= 4 && CV_VERSION_MINOR >= 8
+#define simdSize(type) cv::VTraits<type>::nlanes
+#else
+#define simdSize(type) type::nlanes
+#endif

sum.cpp

@@ -0,0 +1,240 @@
+#include "sum.h"
+
+#include <opencv2/core/hal/intrin.hpp>
+
+inline cv::v_uint32x4 v_add_expand(const cv::v_uint16x8 &src) {
+    cv::v_uint32x4 low;
+    cv::v_uint32x4 high;
+    cv::v_expand(src, low, high);
+
+    return cv::v_add(low, high);
+}
+
+inline cv::v_uint64x2 v_add_expand(const cv::v_uint32x4 &src) {
+    cv::v_uint64x2 low;
+    cv::v_uint64x2 high;
+    cv::v_expand(src, low, high);
+
+    return cv::v_add(low, high);
+}
+
+inline void computeSum(const cv::v_uint8x16 &src, cv::v_uint32x4 &sum, cv::v_uint64x2 &sqSum) {
+    cv::v_uint16x8 low;
+    cv::v_uint16x8 high;
+    cv::v_expand(src, low, high);
+
+    sum            = cv::v_add(sum, v_add_expand(cv::v_add(low, high)));
+    const auto dot = cv::v_dotprod_expand_fast(src, src);
+    sqSum          = cv::v_add(sqSum, v_add_expand(dot));
+}
+
+void computeSum(const cv::Mat &src, const HRegion &hRegion, uint64 &sum, uint64 &sqSum) {
+    constexpr auto blockSize = simdSize(cv::v_uint8);
+    const auto    *srcPtr    = src.data;
+    cv::v_uint32x4 vSum      = cv::v_setzero_u32();
+    cv::v_uint64x2 vSqSum    = cv::v_setzero_u64();
+    uint32_t       partSum   = 0;
+    uint64         partSqSum = 0;
+
+    for (const auto &rle : hRegion) {
+        const auto *ptr = srcPtr + src.step * rle.row + rle.startColumn;
+
+        int i = 0;
+        for (; i < rle.length - blockSize; i += blockSize) {
+            computeSum(cv::v_load(ptr + i), vSum, vSqSum);
+        }
+
+        // TODO aligned fill 0
+        for (; i < rle.length; i++) {
+            const auto val  = ptr[ i ];
+            partSum        += val;
+            partSqSum      += static_cast<ushort>(val) * static_cast<ushort>(val);
+        }
+    }
+
+    sum   = cv::v_reduce_sum(vSum) + partSum;
+    sqSum = cv::v_reduce_sum(vSqSum) + partSqSum;
+}
+
+inline void computeSumDiff(const cv::v_uint16x8 &start, const cv::v_uint16x8 &end,
+                           cv::v_int32x4 &diff0, cv::v_int32x4 &diff1) {
+    cv::v_int16x8 sub;
+    {
+        const auto vStart = cv::v_reinterpret_as_s16(start);
+        const auto vEnd   = cv::v_reinterpret_as_s16(end);
+        sub               = cv::v_sub(vEnd, vStart);
+    }
+
+    cv::v_int32x4 val = cv::v_expand_low(sub);
+    diff0             = cv::v_add(diff0, val);
+
+    val   = cv::v_expand_high(sub);
+    diff1 = cv::v_add(diff1, val);
+}
+
+inline void computeSumDiff(const cv::v_uint8x16 &start, const cv::v_uint8x16 &end,
+                           cv::v_int32x4 &diff0, cv::v_int32x4 &diff1, cv::v_int32x4 &diff2,
+                           cv::v_int32x4 &diff3) {
+    computeSumDiff(cv::v_expand_low(start), cv::v_expand_low(end), diff0, diff1);
+    computeSumDiff(cv::v_expand_high(start), cv::v_expand_high(end), diff2, diff3);
+}
+
+inline void computeSqSumDiff(const cv::v_uint32x4 &start, const cv::v_uint32x4 &end,
+                             cv::v_int32x4 &diff0) {
+    const cv::v_int32x4 vStart = cv::v_reinterpret_as_s32(start);
+    const cv::v_int32x4 vEnd   = cv::v_reinterpret_as_s32(end);
+
+    const cv::v_int32x4 sub = cv::v_sub(vEnd, vStart);
+    diff0                   = cv::v_add(diff0, sub);
+}
+
+inline void computeSqSumDiff(cv::v_uint16x8 &start, cv::v_uint16x8 &end, cv::v_int32x4 &diff0,
+                             cv::v_int32x4 &diff1) {
+    start = cv::v_mul(start, start);
+    end   = cv::v_mul(end, end);
+
+    computeSqSumDiff(cv::v_expand_low(start), cv::v_expand_low(end), diff0);
+    computeSqSumDiff(cv::v_expand_high(start), cv::v_expand_high(end), diff1);
+}
+
+inline void computeSqSumDiff(const cv::v_uint8x16 &start, const cv::v_uint8x16 &end,
+                             cv::v_int32x4 &diff0, cv::v_int32x4 &diff1, cv::v_int32x4 &diff2,
+                             cv::v_int32x4 &diff3) {
+
+    auto vStart = cv::v_expand_low(start);
+    auto vEnd   = cv::v_expand_low(end);
+    computeSqSumDiff(vStart, vEnd, diff0, diff1);
+
+    vStart = cv::v_expand_high(start);
+    vEnd   = cv::v_expand_high(end);
+    computeSqSumDiff(vStart, vEnd, diff2, diff3);
+}
+
+inline void v_expand_store(double *ptr, const std::array<int, 4> &val) {
+    ptr[ 0 ] = ptr[ -1 ] + val[ 0 ];
+    ptr[ 1 ] = ptr[ 0 ] + val[ 1 ];
+    ptr[ 2 ] = ptr[ 1 ] + val[ 2 ];
+    ptr[ 3 ] = ptr[ 2 ] + val[ 3 ];
+}
+
+void shiftH(const uchar *src, std::size_t srcStep, const HRegion &hRegion, int row, double *sum,
+            std::size_t sumStep, int sumWidth, double *sqSum, std::size_t sqSumStep) {
+    constexpr auto blockSize = simdSize(cv::v_uint8);
+    auto          *srcPtr    = src;
+    auto          *sumPtr    = sum + row * sumStep;
+    auto          *sqSumPtr  = sqSum + row * sqSumStep;
+
+    std::array<int, 4> buf{};
+
+    int i = 1;
+    for (; i < sumWidth - blockSize; i += blockSize) {
+        cv::v_int32x4 diff0 = cv::v_setzero_s32();
+        cv::v_int32x4 diff1 = cv::v_setzero_s32();
+        cv::v_int32x4 diff2 = cv::v_setzero_s32();
+        cv::v_int32x4 diff3 = cv::v_setzero_s32();
+
+        cv::v_int32x4 diff10 = cv::v_setzero_s32();
+        cv::v_int32x4 diff11 = cv::v_setzero_s32();
+        cv::v_int32x4 diff12 = cv::v_setzero_s32();
+        cv::v_int32x4 diff13 = cv::v_setzero_s32();
+
+        for (const auto &rle : hRegion) {
+            auto *startPtr = srcPtr + (row + rle.row) * srcStep + rle.startColumn + i - 1;
+            auto *endPtr   = startPtr + rle.length;
+
+            auto vStart = cv::v_load(startPtr);
+            auto vEnd   = cv::v_load(endPtr);
+            computeSumDiff(vStart, vEnd, diff0, diff1, diff2, diff3);
+            computeSqSumDiff(vStart, vEnd, diff10, diff11, diff12, diff13);
+        }
+
+        auto *sumPtrStart = sumPtr + i;
+        cv::v_store(buf.data(), diff0);
+        v_expand_store(sumPtrStart, buf);
+        cv::v_store(buf.data(), diff1);
+        v_expand_store(sumPtrStart + 4, buf);
+        cv::v_store(buf.data(), diff2);
+        v_expand_store(sumPtrStart + 8, buf);
+        cv::v_store(buf.data(), diff3);
+        v_expand_store(sumPtrStart + 12, buf);
+
+        auto *sqSumPtrStart = sqSumPtr + i;
+        cv::v_store(buf.data(), diff10);
+        v_expand_store(sqSumPtrStart, buf);
+        cv::v_store(buf.data(), diff11);
+        v_expand_store(sqSumPtrStart + 4, buf);
+        cv::v_store(buf.data(), diff12);
+        v_expand_store(sqSumPtrStart + 8, buf);
+        cv::v_store(buf.data(), diff13);
+        v_expand_store(sqSumPtrStart + 12, buf);
+    }
+
+    for (; i < sumWidth; i++) {
+        int32_t partSum   = 0;
+        int32_t partSqSum = 0;
+        for (const auto &rle : hRegion) {
+            auto *startPtr = srcPtr + (row + rle.row) * srcStep + rle.startColumn + i - 1;
+            auto *endPtr   = startPtr + rle.length;
+
+            const int32_t start  = *startPtr;
+            const int32_t end    = *endPtr;
+            partSum             += end - start;
+            partSqSum           += end * end - start * start;
+        }
+
+        auto *sumPtrStart   = sumPtr + i;
+        sumPtrStart[ 0 ]    = sumPtrStart[ -1 ] + partSum;
+        auto *sqSumPtrStart = sqSumPtr + i;
+        sqSumPtrStart[ 0 ]  = sqSumPtrStart[ -1 ] + partSqSum;
+    }
+}
+
+void shiftV(const uchar *src, std::size_t srcStep, const VRegion &vRegion, int row, double *sum,
+            std::size_t sumStep, double *sqSum, std::size_t sqSumStep) {
+    auto *srcPtr   = src;
+    auto *sumPtr   = sum + row * sumStep;
+    auto *sqSumPtr = sqSum + row * sqSumStep;
+
+    int32_t partSum   = 0;
+    int32_t partSqSum = 0;
+    for (const auto &rle : vRegion) {
+        auto *startPtr = srcPtr + (row + rle.startRow - 1) * srcStep + rle.col;
+        auto *endPtr   = startPtr + rle.length * srcStep;
+
+        const int32_t start = *startPtr;
+        const int32_t end   = *endPtr;
+
+        partSum   += end - start;
+        partSqSum += end * end - start * start;
+    }
+
+    sumPtr[ 0 ]   = *(sumPtr - sumStep) + partSum;
+    sqSumPtr[ 0 ] = *(sqSumPtr - sqSumStep) + partSqSum;
+}
+
+void integralSum(const cv::Mat &src, cv::Mat &sum, cv::Mat &sqSum, const cv::Size &templateSize,
+                 const HRegion &hRegion, const VRegion &vRegion) {
+    const auto size = src.size() - templateSize + cv::Size(1, 1);
+    sum.create(size, CV_64FC1);
+    sqSum.create(size, CV_64FC1);
+
+    const auto *srcPtr    = src.data;
+    auto       *sumPtr    = reinterpret_cast<double *>(sum.data);
+    auto       *sqSumPtr  = reinterpret_cast<double *>(sqSum.data);
+    const auto  sumStep   = sum.step1();
+    const auto  sqSumStep = sqSum.step1();
+
+    // compute first
+    uint64 sum0;
+    uint64 sqSum0;
+    computeSum(src, hRegion, sum0, sqSum0);
+    sumPtr[ 0 ]   = static_cast<double>(sum0);
+    sqSumPtr[ 0 ] = static_cast<double>(sqSum0);
+
+    for (int y = 0; y < size.height; y++) {
+        shiftH(srcPtr, src.step, hRegion, y, sumPtr, sumStep, sum.cols, sqSumPtr, sqSumStep);
+        if (y + 1 < size.height) {
+            shiftV(srcPtr, src.step, vRegion, y + 1, sumPtr, sumStep, sqSumPtr, sqSumStep);
+        }
+    }
+}

sum.h

@@ -0,0 +1,8 @@
+#pragma once
+
+#include "privateType.h"
+
+#include <opencv2/opencv.hpp>
+
+void integralSum(const cv::Mat &src, cv::Mat &sum, cv::Mat &sqSum, const cv::Size &templateSize,
+                 const HRegion &hRegion, const VRegion &vRegion);