Huawei_Technology
/
mindspore
Not watched
1
0
Fork 0
Code
Releases 13 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
10374 Commits
11 Branches
1.3 GB
177 Download
 
 
 
 
 
 
Tree: 4ff4c17632
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '4ff4c17632'
${ noResults }
mindspore/model_zoo/official/lite/object_detection
gongdaguo 6512f678d5 Fix security error and change download link 5 years ago
..
app Fix security error and change download link 5 years ago
gradle/wrapper add object detection 5 years ago
images add object detection 5 years ago
.gitignore add object detection 5 years ago
README.en.md Fix security error and change download link 5 years ago
README.md Fix security error and change download link 5 years ago
build.gradle add object detection 5 years ago
gradle.properties add object detection 5 years ago
gradlew add object detection 5 years ago
gradlew.bat add object detection 5 years ago
settings.gradle add object detection 5 years ago

README.md

MindSpore Lite 端侧目标检测demo（Android）

本示例程序演示了如何在端侧利用MindSpore Lite C++ API（Android JNI）以及MindSpore Lite 目标检测模型完成端侧推理，实现对图库或者设备摄像头捕获的内容进行检测，并在App图像预览界面中显示连续目标检测结果。

运行依赖

  • Android Studio >= 3.2 (推荐4.0以上版本)
  • NDK 21.3
  • CMake 3.10
  • Android SDK >= 26

构建与运行

  1. 在Android Studio中加载本示例源码，并安装相应的SDK（指定SDK版本后，由Android Studio自动安装）。

    start_home

    启动Android Studio后，点击File->Settings->System Settings->Android SDK，勾选相应的SDK。如下图所示，勾选后，点击OK，Android Studio即可自动安装SDK。

    start_sdk

    （可选）若安装时出现NDK版本问题，可手动下载相应的NDK版本（本示例代码使用的NDK版本为21.3），并在Project Structure的Android NDK location设置中指定SDK的位置。

    project_structure

  2. 连接Android设备，运行目标检测示例应用程序。

    通过USB连接Android设备调试，点击Run 'app'即可在你的设备上运行本示例项目。

    • 注：编译过程中Android Studio会自动下载MindSpore Lite、模型文件等相关依赖项，编译过程需做耐心等待。

    run_app

    Android Studio连接设备调试操作，可参考https://developer.android.com/studio/run/device?hl=zh-cn。

  3. 在Android设备上，点击“继续安装”，安装完即可查看到设备摄像头捕获的内容和推理结果。

    install

    如下图所示，检测出图中内容是鼠标。

    result

示例程序详细说明

本端侧目标检测Android示例程序分为JAVA层和JNI层，其中，JAVA层主要通过Android Camera 2 API实现摄像头获取图像帧，以及相应的图像处理（针对推理结果画框）等功能；JNI层在Runtime中完成模型推理的过程。

此处详细说明示例程序的JNI层实现，JAVA层运用Android Camera 2 API实现开启设备摄像头以及图像帧处理等功能，需读者具备一定的Android开发基础知识。

示例程序结构

app
|
├── libs # 存放demo jni层编译出的库文件
│   └── arm64-v8a
│       │── libmlkit-label-MS.so #
|
├── src/main
│   ├── assets # 资源文件
|   |   └── ssd.ms # 存放模型文件
│   |
│   ├── cpp # 模型加载和预测主要逻辑封装类
|   |   ├── mindspore-lite-x.x.x-mindata-arm64-cpu # minspore源码编译出的调用包,包含demo jni层依赖的库文件及相关的头文件
|   |   |   └── ...
│   |   |
|   |   ├── MindSporeNetnative.cpp # MindSpore调用相关的JNI方法
│   ├── java # java层应用代码
│   │   └── com.huawei.himindsporedemo 
│   │       ├── help # 图像处理及MindSpore JNI调用相关实现
│   │       │   └── ...
│   │       └── obejctdetect # 开启摄像头及绘制相关实现
│   │           └── ...
│   │   
│   ├── res # 存放Android相关的资源文件
│   └── AndroidManifest.xml # Android配置文件
│
├── CMakeList.txt # cmake编译入口文件
│
├── build.gradle # 其他Android配置文件
├── download.gradle # APP构建时由gradle自动从HuaWei Server下载依赖的库文件及模型文件
└── ...

配置MindSpore Lite依赖项

Android JNI层调用MindSpore C++ API时，需要相关库文件支持。可通过MindSpore Lite源码编译生成"mindspore-lite-X.X.X-mindata-armXX-cpu"库文件包（包含libmindspore-lite.so库文件和相关头文件，可包含多个兼容架构）。

在Android Studio中将编译完成的mindspore-lite-X.X.X-mindata-armXX-cpu压缩包，解压之后放置在APP工程的app/src/main/cpp目录下，并在app的build.gradle文件中配置CMake编译支持，以及arm64-v8a和armeabi-v7a的编译支持，如下所示：

android{
    defaultConfig{
        externalNativeBuild{
            cmake{
                arguments "-DANDROID_STL=c++_shared"
            }
        }

        ndk{ 
            abiFilters 'arm64-v8a'
        }
    }
}

在app/CMakeLists.txt文件中建立.so库文件链接，如下所示。

# Set MindSpore Lite Dependencies.
set(MINDSPORELITE_VERSION  mindspore-lite-1.0.0-minddata-arm64-cpu)
include_directories(${CMAKE_SOURCE_DIR}/src/main/cpp/${MINDSPORELITE_VERSION})
add_library(mindspore-lite SHARED IMPORTED )
add_library(minddata-lite SHARED IMPORTED )
set_target_properties(mindspore-lite PROPERTIES IMPORTED_LOCATION
        ${CMAKE_SOURCE_DIR}/src/main/cpp/${MINDSPORELITE_VERSION}/lib/libmindspore-lite.so)
set_target_properties(minddata-lite PROPERTIES IMPORTED_LOCATION
        ${CMAKE_SOURCE_DIR}/src/main/cpp/${MINDSPORELITE_VERSION}/lib/libminddata-lite.so)

# Link target library.       
target_link_libraries(
    ...
    mindspore-lite
    minddata-lite
    ...
)

本示例中，app build过程由download.gradle文件自动从华为服务器下载mindspore所编译的库及相关头文件，并放置在src/main/cpp工程目录下。

  • 注：若自动下载失败，请手动下载相关库文件并将其放在对应位置：
  • mindspore-lite-1.0.0-minddata-arm64-cpu.tar.gz 下载链接

下载及部署模型文件

从MindSpore Model Hub中下载模型文件，本示例程序中使用的目标检测模型文件为ssd.ms，同样通过download.gradle脚本在APP构建时自动下载，并放置在app/src/main/assets工程目录下。

  • 注：若下载失败请手动下载模型文件，ssd.ms 下载链接。

编写端侧推理代码

在JNI层调用MindSpore Lite C++ API实现端测推理。

推理代码流程如下，完整代码请参见src/cpp/MindSporeNetnative.cpp。

  1. 加载MindSpore Lite模型文件，构建上下文、会话以及用于推理的计算图。

    • 加载模型文件：创建并配置用于模型推理的上下文

      // Buffer is the model data passed in by the Java layer
jlong bufferLen = env->GetDirectBufferCapacity(buffer);
char *modelBuffer = CreateLocalModelBuffer(env, buffer);

    • 创建会话

      void **labelEnv = new void *;
MSNetWork *labelNet = new MSNetWork;
*labelEnv = labelNet;

// Create context.
lite::Context *context = new lite::Context;
context->cpu_bind_mode_ = lite::NO_BIND;
context->device_ctx_.type = lite::DT_CPU;
context->thread_num_ = numThread;  //Specify the number of threads to run inference

// Create the mindspore session.
labelNet->CreateSessionMS(modelBuffer, bufferLen, "device label", context);
delete context;

    • 加载模型文件并构建用于推理的计算图

      void MSNetWork::CreateSessionMS(char* modelBuffer, size_t bufferLen, std::string name, mindspore::lite::Context* ctx)
{
    CreateSession(modelBuffer, bufferLen, ctx);  
    session = mindspore::session::LiteSession::CreateSession(ctx);
    auto model = mindspore::lite::Model::Import(modelBuffer, bufferLen);
    int ret = session->CompileGraph(model); // Compile Graph 
}

  2. 将输入图片转换为传入MindSpore模型的Tensor格式。

    将待检测图片数据转换为输入MindSpore模型的Tensor。

    // Convert the Bitmap image passed in from the JAVA layer to Mat for OpenCV processing
    LiteMat lite_mat_bgr,lite_norm_mat_cut;

   if (!BitmapToLiteMat(env, srcBitmap, lite_mat_bgr)){
    MS_PRINT("BitmapToLiteMat error");
       return NULL;
   }
   int srcImageWidth = lite_mat_bgr.width_;
   int srcImageHeight = lite_mat_bgr.height_;
   if(!PreProcessImageData(lite_mat_bgr, lite_norm_mat_cut)){
    MS_PRINT("PreProcessImageData error");
       return NULL;
   }
   ImgDims inputDims;
   inputDims.channel =lite_norm_mat_cut.channel_;
   inputDims.width = lite_norm_mat_cut.width_;
   inputDims.height = lite_norm_mat_cut.height_;

   // Get the mindsore inference environment which created in loadModel().
   void **labelEnv = reinterpret_cast<void **>(netEnv);
   if (labelEnv == nullptr) {
       MS_PRINT("MindSpore error, labelEnv is a nullptr.");
       return NULL;
   }
   MSNetWork *labelNet = static_cast<MSNetWork *>(*labelEnv);

   auto mSession = labelNet->session;
   if (mSession == nullptr) {
       MS_PRINT("MindSpore error, Session is a nullptr.");
       return NULL;
   }
   MS_PRINT("MindSpore get session.");

   auto msInputs = mSession->GetInputs();
   auto inTensor = msInputs.front();

   float *dataHWC = reinterpret_cast<float *>(lite_norm_mat_cut.data_ptr_);
   // copy input Tensor
   memcpy(inTensor->MutableData(), dataHWC,
          inputDims.channel * inputDims.width * inputDims.height * sizeof(float));
   delete[] (dataHWC);

  3. 进行模型推理前，输入tensor格式为 NHWC，shape为1:300:300:3，格式为RGB, 并对输入tensor做标准化处理.

     bool PreProcessImageData(LiteMat &lite_mat_bgr,LiteMat &lite_norm_mat_cut) {
    bool ret=false;
    LiteMat lite_mat_resize;
    ret = ResizeBilinear(lite_mat_bgr, lite_mat_resize, 300, 300);
    if (!ret) {
        MS_PRINT("ResizeBilinear error");
        return false;
    }
    LiteMat lite_mat_convert_float;
    ret = ConvertTo(lite_mat_resize, lite_mat_convert_float, 1.0 / 255.0);
    if (!ret) {
        MS_PRINT("ConvertTo error");
        return false;
    }

    float means[3] = {0.485, 0.456, 0.406};
    float vars[3] = {1.0 / 0.229, 1.0 / 0.224, 1.0 / 0.225};
    SubStractMeanNormalize(lite_mat_convert_float, lite_norm_mat_cut, means, vars);
    return true;
}

  4. 对输入Tensor按照模型进行推理，获取输出Tensor，并进行后处理。

    • 图执行，端测推理。

      // After the model and image tensor data is loaded, run inference.
auto status = mSession->RunGraph();

    • 获取输出数据。

      auto names = mSession->GetOutputTensorNames();
    typedef std::unordered_map<std::string,
            std::vector<mindspore::tensor::MSTensor *>> Msout;
std::unordered_map<std::string,
            mindspore::tensor::MSTensor *> msOutputs;
for (const auto &name : names) {
        auto temp_dat =mSession->GetOutputByTensorName(name);  
        msOutputs.insert(std::pair<std::string, mindspore::tensor::MSTensor *> {name, temp_dat});  
    }
std::string retStr = ProcessRunnetResult(msOutputs, ret);

    • 模型有2个输出，输出1是目标的类别置信度，维度为1：1917: 81； 输出2是目标的矩形框坐标偏移量，维度为1:1917:4。 为了得出目标的实际矩形框，需要根据偏移量计算出矩形框的位置。这部分在 getDefaultBoxes中实现。

      void SSDModelUtil::getDefaultBoxes() {
    float fk[6] = {0.0, 0.0, 0.0, 0.0, 0.0, 0.0};
    std::vector<struct WHBox> all_sizes;
    struct Product mProductData[19 * 19] = {0};

    for (int i = 0; i < 6; i++) {
        fk[i] = config.model_input_height / config.steps[i];
    }
    float scale_rate =
            (config.max_scale - config.min_scale) / (sizeof(config.num_default) / sizeof(int) - 1);
    float scales[7] = {0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0};
    for (int i = 0; i < sizeof(config.num_default) / sizeof(int); i++) {
        scales[i] = config.min_scale + scale_rate * i;
    }

    for (int idex = 0; idex < sizeof(config.feature_size) / sizeof(int); idex++) {
        float sk1 = scales[idex];
        float sk2 = scales[idex + 1];
        float sk3 = sqrt(sk1 * sk2);
        struct WHBox tempWHBox;

        all_sizes.clear();

        if (idex == 0) {
            float w = sk1 * sqrt(2);
            float h = sk1 / sqrt(2);

            tempWHBox.boxw = 0.1;
            tempWHBox.boxh = 0.1;
            all_sizes.push_back(tempWHBox);

            tempWHBox.boxw = w;
            tempWHBox.boxh = h;
            all_sizes.push_back(tempWHBox);

            tempWHBox.boxw = h;
            tempWHBox.boxh = w;
            all_sizes.push_back(tempWHBox);

        } else { 
            tempWHBox.boxw = sk1;
            tempWHBox.boxh = sk1;
            all_sizes.push_back(tempWHBox);

            for (int j = 0; j < sizeof(config.aspect_ratios[idex]) / sizeof(int); j++) {
                float w = sk1 * sqrt(config.aspect_ratios[idex][j]);
                float h = sk1 / sqrt(config.aspect_ratios[idex][j]);
                tempWHBox.boxw = w;
                tempWHBox.boxh = h;
                all_sizes.push_back(tempWHBox);
                tempWHBox.boxw = h;
                tempWHBox.boxh = w;
                all_sizes.push_back(tempWHBox);
            }

            tempWHBox.boxw = sk3;
            tempWHBox.boxh = sk3;
            all_sizes.push_back(tempWHBox);
        }

        for (int i = 0; i < config.feature_size[idex]; i++) {
            for (int j = 0; j < config.feature_size[idex]; j++) {
                mProductData[i * config.feature_size[idex] + j].x = i;
                mProductData[i * config.feature_size[idex] + j].y = j;
            }
        }

        int productLen = config.feature_size[idex] * config.feature_size[idex];

        for (int i = 0; i < productLen; i++) {
            for (int j = 0; j < all_sizes.size(); j++) {
                struct NormalBox tempBox;
                float cx = (mProductData[i].y + 0.5) / fk[idex];
                float cy = (mProductData[i].x + 0.5) / fk[idex];
                tempBox.y = cy;
                tempBox.x = cx;
             tempBox.h = all_sizes[j].boxh;
                tempBox.w = all_sizes[j].boxw;
             mDefaultBoxes.push_back(tempBox);
            }
        }
    }
}

    • 通过最大值抑制将目标类型置信度较高的输出筛选出来。

      void SSDModelUtil::nonMaximumSuppression(const YXBoxes *const decoded_boxes,
                                         const float *const scores,
                                         const std::vector<int> &in_indexes,
                                         std::vector<int> &out_indexes, const float nmsThreshold,
                                         const int count, const int max_results) {
    int nR = 0; //number of results
    std::vector<bool> del(count, false);
    for (size_t i = 0; i < in_indexes.size(); i++) {
        if (!del[in_indexes[i]]) {
            out_indexes.push_back(in_indexes[i]);
            if (++nR == max_results) {
                break;
            }
            for (size_t j = i + 1; j < in_indexes.size(); j++) {
                const auto boxi = decoded_boxes[in_indexes[i]], boxj = decoded_boxes[in_indexes[j]];
                float a[4] = {boxi.xmin, boxi.ymin, boxi.xmax, boxi.ymax};
                float b[4] = {boxj.xmin, boxj.ymin, boxj.xmax, boxj.ymax};
             if (IOU(a, b) > nmsThreshold) {
                    del[in_indexes[j]] = true;
                }
            }
        }
    }
}

    • 对每类的概率大于阈值，通过NMS算法筛选出矩形框后， 还需要将输出的矩形框恢复到原图尺寸。

      std::string SSDModelUtil::getDecodeResult(float *branchScores, float *branchBoxData) {
    std::string result = "";
    NormalBox tmpBox[1917] = {0};
    float mScores[1917][81] = {0};
    float outBuff[1917][7] = {0};
    float scoreWithOneClass[1917] = {0};
    int outBoxNum = 0;
    YXBoxes decodedBoxes[1917] = {0};

    // Copy branch outputs box data to tmpBox.
    for (int i = 0; i < 1917; ++i) {
        tmpBox[i].y = branchBoxData[i * 4 + 0];
        tmpBox[i].x = branchBoxData[i * 4 + 1];
        tmpBox[i].h = branchBoxData[i * 4 + 2];
        tmpBox[i].w = branchBoxData[i * 4 + 3];
    }

    // Copy branch outputs score to mScores.
    for (int i = 0; i < 1917; ++i) {
        for (int j = 0; j < 81; ++j) {
            mScores[i][j] = branchScores[i * 81 + j];
        }
    }

     ssd_boxes_decode(tmpBox, decodedBoxes);
     const float nms_threshold = 0.3;
     for (int i = 1; i < 81; i++) {
         std::vector<int> in_indexes;
         for (int j = 0; j < 1917; j++) {
             scoreWithOneClass[j] = mScores[j][i];
          //   if (mScores[j][i] > 0.1) {
             if (mScores[j][i] > g_thres_map[i]) {
                 in_indexes.push_back(j);
             }
         }
         if (in_indexes.size() == 0) {
             continue;
         }

         sort(in_indexes.begin(), in_indexes.end(),
              [&](int a, int b) { return scoreWithOneClass[a] > scoreWithOneClass[b]; });
         std::vector<int> out_indexes;

         nonMaximumSuppression(decodedBoxes, scoreWithOneClass, in_indexes, out_indexes,
                               nms_threshold);
         for (int k = 0; k < out_indexes.size(); k++) {
             outBuff[outBoxNum][0] = out_indexes[k]; //image id
             outBuff[outBoxNum][1] = i; //labelid
             outBuff[outBoxNum][2] = scoreWithOneClass[out_indexes[k]]; //scores
             outBuff[outBoxNum][3] =
                     decodedBoxes[out_indexes[k]].xmin * inputImageWidth / 300;
             outBuff[outBoxNum][4] =
                     decodedBoxes[out_indexes[k]].ymin * inputImageHeight / 300;
             outBuff[outBoxNum][5] =
                     decodedBoxes[out_indexes[k]].xmax * inputImageWidth / 300;
             outBuff[outBoxNum][6] =
                     decodedBoxes[out_indexes[k]].ymax * inputImageHeight / 300;
             outBoxNum++;
         }
     }
     MS_PRINT("outBoxNum %d", outBoxNum);

     for (int i = 0; i < outBoxNum; ++i) {
         std::string tmpid_str = std::to_string(outBuff[i][0]);
         result += tmpid_str; // image ID
         result += "_";
         // tmpid_str = std::to_string(outBuff[i][1]);
         MS_PRINT("label_classes i %d, outBuff %d",i, (int) outBuff[i][1]);
         tmpid_str = label_classes[(int) outBuff[i][1]];
         result += tmpid_str; // label id
         result += "_";
         tmpid_str = std::to_string(outBuff[i][2]);
         result += tmpid_str; // scores
         result += "_";
         tmpid_str = std::to_string(outBuff[i][3]);
         result += tmpid_str; // xmin
         result += "_";
         tmpid_str = std::to_string(outBuff[i][4]);
         result += tmpid_str; // ymin
         result += "_";
         tmpid_str = std::to_string(outBuff[i][5]);
         result += tmpid_str; // xmax
         result += "_";
         tmpid_str = std::to_string(outBuff[i][6]);
         result += tmpid_str; // ymax
         result += ";";
     }
    return result;
}

​

Delete task

Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.

cancel
confirm

MindSpore is a new open source deep learning training/inference framework that could be used for mobile, edge and cloud scenarios.

C++ Python Text Unity3D Asset C other