NCNN----Monodepthv2单目深度估计小米手机部署-CSDN博客

本文链接：https://blog.csdn.net/qq_37172182/article/details/130543661

文章介绍了如何理解和转化MonoDepthv2模型，将其优化为NCNN格式，以便在小米手机上进行高效的推理。过程中涉及模型源码分析、转换工具如onnx2ncnn的使用，以及AndroidStudio项目配置，包括CMakeLists.txt的修改和OpenCV、ncnn库的集成。最终在Android应用中实现了模型的加载和设备选择功能。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

题目要求：学习了解单目深度估计模型MonoDepthv2，基于NCNN推理框架部署到小米手机
MonoDepthv2 论文：Digging Into Self-Supervised Monocular Depth Estimation
MonoDepthv2 源码：Monodepth2 GitHub

分析：
1）了解MonoDepthv2的基本原理和代码理解
2）将模型转化为更加方便高效的NCNN模型并在小米手机端完成推理过程

结果展示：

在这里插入图片描述

参考代码
nihui/ncnn-android-nanodet
模型转换
一键转换 onnx2ncnn
包依赖
opencv-mobile-4.6.0-android (版本可选)
ncnn-20230223-android-vulkan (版本可选)
软件环境
Android Studio, SDK Platforms: Android 11, 12, 13; SDK tools, NDK tools
工程目录（重点修改部分已经标红）

具体修改项

CMakeLists.txt

project(nanodetncnn)

cmake_minimum_required(VERSION 3.10)

set(OpenCV_DIR ${CMAKE_SOURCE_DIR}/opencv-mobile-4.6.0-android/sdk/native/jni)
find_package(OpenCV REQUIRED core imgproc)

set(ncnn_DIR ${CMAKE_SOURCE_DIR}/ncnn-20230223-android-vulkan/${ANDROID_ABI}/lib/cmake/ncnn)
find_package(ncnn REQUIRED)
add_library(nanodetncnn SHARED nanodetncnn.cpp nanodet.cpp ndkcamera.cpp)

target_link_libraries(nanodetncnn ncnn ${OpenCV_LIBS} camera2ndk mediandk)

nanodet.cpp (#59)设置模型输入输出尺寸

int NanoDet::detect(cv::Mat& rgb)
{
    ncnn::Mat in = ncnn::Mat::from_pixels_resize(rgb.data, ncnn::Mat::PIXEL_RGB, rgb.cols, rgb.rows,640, 192);
    in.substract_mean_normalize(mean_vals, norm_vals);

    ncnn::Extractor ex = nanodet.create_extractor();
    ex.input("input", in);
    ncnn::Mat model_out;
    ex.extract("output", model_out);

    cv::Mat out(192, 640, CV_32FC1, model_out.data);

string.xml

<?xml version="1.0" encoding="utf-8"?>
<resources>
    <string name="app_name">nanodetncnn</string>
    <string-array name="model_array">
        <item>MonoDepthv2</item>
<!--        <item>MobilenetV2-Wave</item>-->
<!--        <item>DenseNet</item>-->
<!--        <item>DenseNet-Wave</item>-->
    </string-array>
    <string-array name="cpugpu_array">
        <item>CPU</item>
        <item>GPU</item>
    </string-array>
</resources>

nanodetncnn.cpp (#170) 设置模型路径和设备cpugpu选项

JNIEXPORT jboolean JNICALL Java_com_tencent_nanodetncnn_NanoDetNcnn_loadModel(JNIEnv* env, jobject thiz, jobject assetManager, jint modelid, jint cpugpu)
{
    // 检查一下选的模型和设备是不是在范围
    if (modelid < 0 || modelid > 2 || cpugpu < 0 || cpugpu > 1)
    {
        return JNI_FALSE;
    }

    AAssetManager* mgr = AAssetManager_fromJava(env, assetManager);

    __android_log_print(ANDROID_LOG_DEBUG, "ncnn", "loadModel %p", mgr);

    const char* modeltypes[] =
    {"mono-sim-opt",
    };

gradle-wrapper.properties

distributionBase=GRADLE_USER_HOME
distributionPath=wrapper/dists
zipStoreBase=GRADLE_USER_HOME
zipStorePath=wrapper/dists
#distributionUrl=https\://services.gradle.org/distributions/gradle-5.4.1-all.zip
distributionUrl=https\://mirrors.cloud.tencent.com/gradle/gradle-5.4.1-all.zip

local.properties

## This file must *NOT* be checked into Version Control Systems,
# as it contains information specific to your local configuration.
#
# Location of the SDK. This is only used by Gradle.
# For customization when using a Version Control System, please read the
# header note.
#Tue Apr 11 21:28:46 CST 2023
sdk.dir=\\path\\to\\AppData\\Local
ndk.dir=\\path\\to\\AppData\\Local\\ndk\\22.1.7171670