于老师您好！

我在用Raspberry Pi 3 B+做调试，看到项目README的结论，这款设备在320x240的图片下，单核性能是8.1fps，多核性能是23.74fps。但是我这边在320x240的实际测试结果，多核只有4.2fps，远远低于前面给出的结论。方便的话，请帮我找找原因，谢谢！

//下面是一些环境参数

1. os: 2020-02-13-raspbian-buster-lite
2. gcc version: 8.3.0 (Raspbian 8.3.0-6+rpi1) (本地编译)
3. cmake options: AVX512 = OFF AVX2 = OFF NEON = ON OpenMP = TRUE DEMO = ON add_compile_options(-mfpu=neon) (不加编译不通过)
4. opencv: 3.2.0

//cmake的输出 -- The C compiler identification is GNU 8.3.0 -- The CXX compiler identification is GNU 8.3.0 -- Check for working C compiler: /usr/bin/cc -- Check for working C compiler: /usr/bin/cc -- works -- Detecting C compiler ABI info -- Detecting C compiler ABI info - done -- Detecting C compile features -- Detecting C compile features - done -- Check for working CXX compiler: /usr/bin/c++ -- Check for working CXX compiler: /usr/bin/c++ -- works -- Detecting CXX compiler ABI info -- Detecting CXX compiler ABI info - done -- Detecting CXX compile features -- Detecting CXX compile features - done fatal: No names found, cannot describe anything. BUILD_VERSION:v0.0.1 Using ENON -- Found OpenMP_C: -fopenmp (found version "4.5") -- Found OpenMP_CXX: -fopenmp (found version "4.5") -- Found OpenMP: TRUE (found version "4.5")
-- Performing Test COMPILER_HAS_HIDDEN_VISIBILITY -- Performing Test COMPILER_HAS_HIDDEN_VISIBILITY - Success -- Performing Test COMPILER_HAS_HIDDEN_INLINE_VISIBILITY -- Performing Test COMPILER_HAS_HIDDEN_INLINE_VISIBILITY - Success -- Performing Test COMPILER_HAS_DEPRECATED_ATTR -- Performing Test COMPILER_HAS_DEPRECATED_ATTR - Success -- Found OpenCV: /usr (found version "3.2.0") AVX512 = OFF AVX2 = OFF NEON = ON OpenMP = TRUE DEMO = ON -- Configuring done -- Generating done -- Build files have been written to: /home/pi/dev/libfacedetection/build

//benchmark的运行结果 [email protected]:~/dev/libfacedetection/build $ ./benchmark 320_0.jpg There are 4 threads, 4 processors. cnn facedetection average time = 238.36ms | 4.20 FPS

于老师，您好！

我在windows系统运行图片人脸检测，912*524分辨率的png图片，可以正常完整运行，但是facedetect_cnn 这个函数耗时1903ms，我知道这是不合理的，但是我没有找到原因。电脑配置不差。使用默认的cmakeLists编译的facedetection_export.h，使用detect-image.cpp的示例代码做了一点修改，把图片地址写在代码里的。请问是会是什么原因导致？

Do you have any activity for using your algorithm for detecting face on thermal images? It is very interesting now especially by Corona, where automatical facial temperature detection

您好，我用VS2013在x86和x64环境下没有错误提示，但运行示例时libfacedetectcnn-example都出现以下错误提示，请大家帮忙指点： 一直错误提示： 0x000000013F659557 处的第一机会异常(在 LibFaceDetection.exe 中): 0xC0000005: 读取位置 0xFFFFFFFFFFFFFFFF 时发生访问冲突。

代码停留在： _mm256_store_si256((__m256i*)sumarray, sum_int32x8);

没有更改过源文件。

I ported libfacedetection with jni and build with ndk19, then ran on a Nokia 7 plus android phone which cpu is Qualcomm 660. The face rect is given correctly but took almost 700ms to get the result without couting the convert time of bitmap to rgb. Barely can it run 2fps... And last time I check ,NEON was enabled. Can anyone help me with it ?

I use your library in java with jni. And this error happens if i'm run "facedetect_frontal" function about 350-400 times. Can you tell me what mean this error ? How i can resolve it ? Thank you. ` #include "faceSwap_OurNativeLib.h" #include "facedetect-dll.h" #include <opencv2/opencv.hpp> #include #include #include using namespace std; using namespace cv; #define DETECT_BUFFER_SIZE 0x20000

JNIEXPORT jintArray JNICALL Java_faceswap_OurNativeLib_getPointsAndRec (JNIEnv * env, jobject obj, jlong matAddr, jfloat scale, jint min_neighbors,jint minObjectWidth,jint maxObjectWidth,jint dolandmark) { jintArray rezult = nullptr; Mat * inputMat = (Mat*)matAddr; Mat gray; cvtColor(*inputMat, gray, CV_BGR2GRAY); int * pResults = NULL; //pBuffer is used in the detection functions. //If you call functions in multiple threads, please create one buffer for each thread! unsigned char * pBuffer = (unsigned char *)malloc(DETECT_BUFFER_SIZE); if (!pBuffer) { fprintf(stderr, "Can not alloc buffer.\n"); rezult = env->NewIntArray(1); return rezult; }

///////////////////////////////////////////
// frontal face detection / 68 landmark detection
// it's fast, but cannot detect side view faces
//////////////////////////////////////////
//!!! The input image must be a gray one (single-channel)
//!!! DO NOT RELEASE pResults !!!
pResults = facedetect_frontal(pBuffer, (unsigned char*)(gray.ptr(0)), gray.cols, gray.rows, (int)gray.step,
	scale, min_neighbors, minObjectWidth, maxObjectWidth, dolandmark);


if ((*pResults) <= 0)
{
	fprintf(stderr, "Nofaces.\n");
	rezult = env->NewIntArray(1);
	return rezult;
}

rezult = env->NewIntArray((*pResults) * 140);
if (rezult == nullptr)
{
	fprintf(stderr, "Can not alloc rezult memory.\n");
	rezult = env->NewIntArray(1);
	return rezult;
}
//jint bodyOfrezult[140];
//for (int i = 0; i < (pResults ? *pResults : 0); i++)
//{
//	short * p = ((short*)(pResults + 1)) + 142 * i;
//	for (int k = 0; k < 4; k++)
//	{
//		bodyOfrezult[k] = p[k];
//	}
//	if (dolandmark)
//	{
//		for (int j = 4; j < 140; j++)
//		{
//			bodyOfrezult[j] = ((int)p[2 + j]);
//		}
//		env->SetIntArrayRegion(rezult, i * 140, 140 , bodyOfrezult);
//	}
//}
free(pBuffer);
gray.release();
return rezult;

}

`

This lib is very useful! I try it in iOS and successful run.

1. download or clone this lib in your computer;
2. create New Xcode project;
3. add this lib's src to your project;
4. add the system lib libc++.tbd and other you need framework(eg. ACFoundation.framework etc.);
5. download opencv2.framework and add it in your project;
6. follow the example file to write the code.

!!!

1. modify facedetectcnn.h

//#define _ENABLE_AVX2 //Please enable it if X64 CPU
#define _ENABLE_NEON //Please enable it if ARM CPU

2. modify .m to .mm
3. import lib in your .mm

#import <opencv2/opencv.hpp>
#import <opencv2/imgcodecs/ios.h>
#import "ViewController.h"
#import "facedetectcnn.h"

you must import the opencv2/opencv.hpp first !!!

MyCode:

#import <opencv2/opencv.hpp>
#import <opencv2/imgcodecs/ios.h>
#import "ViewController.h"
#import "facedetectcnn.h"

//define the buffer size. Do not change the size!
#define DETECT_BUFFER_SIZE 0x20000
using namespace cv;

@implementation ViewController

- (UIImage *)loadImageAndDectect:(const char *)image_file{
    Mat img = imread(image_file);
    if (img.empty()) {
        fprintf(stderr, "Can not load the image file %s.\n", image_file);
        return nil;
    }
    
    int *pResults = NULL;
    unsigned char * pBuffer = (unsigned char *)malloc(DETECT_BUFFER_SIZE);
    if (!pBuffer) {
        fprintf(stderr, "Can not alloc buffer.\n");
        return nil;
    }
    pResults = facedetect_cnn(pBuffer, (unsigned char *)(img.ptr(0)), img.cols, img.rows, (int)img.step);
    printf("%d faces detected.\n", (pResults ? *pResults : 0));
    Mat result_cnn = img.clone();;
    //print the detection results
    for(int i = 0; i < (pResults ? *pResults : 0); i++)
    {
        short * p = ((short*)(pResults+1))+142*i;
        int x = p[0];
        int y = p[1];
        int w = p[2];
        int h = p[3];
        int neighbors = p[4];
        int angle = p[5];
        
        printf("face_rect=[%d, %d, %d, %d], neighbors=%d, angle=%d\n", x,y,w,h,neighbors, angle);
        rectangle(result_cnn, cv::Rect(x, y, w, h), Scalar(0, 255, 0), 2);
    }
    
    free(pBuffer);
    
    return MatToUIImage(result_cnn);
}


- (void)viewDidLoad {
    [super viewDidLoad];
    // Do any additional setup after loading the view, typically from a nib.
    
    UIImageView *imageView = [[UIImageView alloc] initWithFrame:[UIScreen mainScreen].bounds];
    imageView.contentMode = UIViewContentModeScaleAspectFit;
    [self.view addSubview:imageView];
    
    NSString *path = [[NSBundle mainBundle] pathForResource:@"test" ofType:@".jpg"];
    
    imageView.image = [self loadImageAndDectect:[path UTF8String]];
}


@end

[email protected]:~/Projects/libfacedetection/build$ make -j4
[ 16%] Building CXX object CMakeFiles/demo.dir/src/facedetectcnn.cpp.o
/home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp: In function ‘float dotProductFloatChGeneral(float*, float*, int, int)’:
/home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:115:39: warning: AVX vector return without AVX enabled changes the ABI [-Wpsabi]
     __m256 sumvec = _mm256_setzero_ps();
                                       ^
In file included from /usr/lib/gcc/x86_64-linux-gnu/7/include/immintrin.h:41:0,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.h:54,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:39:
/usr/lib/gcc/x86_64-linux-gnu/7/include/avxintrin.h:1227:1: error: inlining failed in call to always_inline ‘__m256 _mm256_setzero_ps()’: target specific option mismatch
 _mm256_setzero_ps (void)
 ^~~~~~~~~~~~~~~~~
/home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:115:38: note: called from here
     __m256 sumvec = _mm256_setzero_ps();
                     ~~~~~~~~~~~~~~~~~^~
In file included from /usr/lib/gcc/x86_64-linux-gnu/7/include/immintrin.h:85:0,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.h:54,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:39:
/usr/lib/gcc/x86_64-linux-gnu/7/include/fmaintrin.h:63:1: error: inlining failed in call to always_inline ‘__m256 _mm256_fmadd_ps(__m256, __m256, __m256)’: target specific option mismatch
 _mm256_fmadd_ps (__m256 __A, __m256 __B, __m256 __C)
 ^~~~~~~~~~~~~~~
/home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:127:33: note: called from here
         sumvec = _mm256_fmadd_ps(avec, bvec, sumvec);
                  ~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~
In file included from /usr/lib/gcc/x86_64-linux-gnu/7/include/immintrin.h:41:0,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.h:54,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:39:
/usr/lib/gcc/x86_64-linux-gnu/7/include/avxintrin.h:871:1: error: inlining failed in call to always_inline ‘__m256 _mm256_load_ps(const float*)’: target specific option mismatch
 _mm256_load_ps (float const *__P)
 ^~~~~~~~~~~~~~
/home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:120:14: note: called from here
         bvec = _mm256_load_ps(p2 + i);
         ~~~~~^~~~~~~~~~~~~~~~~~~~~~~~
In file included from /usr/lib/gcc/x86_64-linux-gnu/7/include/immintrin.h:41:0,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.h:54,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:39:
/usr/lib/gcc/x86_64-linux-gnu/7/include/avxintrin.h:871:1: error: inlining failed in call to always_inline ‘__m256 _mm256_load_ps(const float*)’: target specific option mismatch
 _mm256_load_ps (float const *__P)
 ^~~~~~~~~~~~~~
/home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:119:14: note: called from here
         avec = _mm256_load_ps(p1 + i);
         ~~~~~^~~~~~~~~~~~~~~~~~~~~~~~
In file included from /usr/lib/gcc/x86_64-linux-gnu/7/include/immintrin.h:41:0,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.h:54,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:39:
/usr/lib/gcc/x86_64-linux-gnu/7/include/avxintrin.h:269:1: error: inlining failed in call to always_inline ‘__m256 _mm256_hadd_ps(__m256, __m256)’: target specific option mismatch
 _mm256_hadd_ps (__m256 __X, __m256 __Y)
 ^~~~~~~~~~~~~~
/home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:131:28: note: called from here
     sumvec = _mm256_hadd_ps(sumvec, sumvec);
              ~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~
In file included from /usr/lib/gcc/x86_64-linux-gnu/7/include/immintrin.h:41:0,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.h:54,
                 from /home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:39:
/usr/lib/gcc/x86_64-linux-gnu/7/include/avxintrin.h:269:1: error: inlining failed in call to always_inline ‘__m256 _mm256_hadd_ps(__m256, __m256)’: target specific option mismatch
 _mm256_hadd_ps (__m256 __X, __m256 __Y)
 ^~~~~~~~~~~~~~
/home/lvmc/Projects/libfacedetection/src/facedetectcnn.cpp:132:28: note: called from here
     sumvec = _mm256_hadd_ps(sumvec, sumvec);
              ~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~
CMakeFiles/demo.dir/build.make:134: recipe for target 'CMakeFiles/demo.dir/src/facedetectcnn.cpp.o' failed
make[2]: *** [CMakeFiles/demo.dir/src/facedetectcnn.cpp.o] Error 1
CMakeFiles/Makefile2:67: recipe for target 'CMakeFiles/demo.dir/all' failed
make[1]: *** [CMakeFiles/demo.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2

我在Ubantu上运行g++ libfacedetectcnn-example.cpp ./test.jpg -O3 pkg-config --libs --cflags opencv4 undefined reference to `facedetect_cnn(unsigned char*, unsigned char*, int, int, int)'， 我已经加头文件了啊

Why is it that after I migrated to Android, the recognition time of keliamoniz1.jpg in the sample image was 2800 milliseconds?

This is my CMakeList.txt:

cmake_minimum_required(VERSION 3.4.1)

SET(CMAKE_BUILD_TYPE Release)

set(CMAKE_SYSTEM_NAME Linux)
set(CMAKE_SYSTEM_VERSION 1)
set(CMAKE_SYSTEM_PROCESSOR "aarch64")

add_definitions(-O3)

add_definitions(-D_ENABLE_INT8)
#add_definitions(-D_ENABLE_AVX2)

SET(CMAKE_C_FLAGS "$ENV{CFLAGS} -O3 -Wall")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=gnu++11 -O3 -pthread")
SET(CMAKE_CXX_FLAGS_RELEASE "$ENV{CXXFLAGS} -O3 -Wall")

#set(CMAKE_C_COMPILER_ABI armeabi-v7a)

set(OPENCV_PATH /home/ww/Downloads/OpenCV-android-sdk)
set(CMAKE_VERBOSE_MAKEFILE on)

include_directories(${OPENCV_PATH}/sdk/native/jni/include)
add_library(lib_opencv STATIC IMPORTED)
set_target_properties(lib_opencv
        PROPERTIES
        IMPORTED_LOCATION ${PROJECT_SOURCE_DIR}/src/main/jniLibs/${ANDROID_ABI}/libopencv_java4.so
        )

add_library( 
        native-lib

        SHARED

        src/main/cpp/native-lib.cpp
        src/main/cpp/facedetectcnn.cpp
        src/main/cpp/facedetectcnn-floatdata.cpp
        src/main/cpp/facedetectcnn-int8data.cpp
        src/main/cpp/facedetectcnn-model.cpp
        )

find_library( 
        log-lib
        log)

target_link_libraries( 
        native-lib
        ${log-lib}
        lib_opencv
        )

Android device CPU is : mtk6737

您好，11 月 10 日更新的这个版本您说的是可以在多线程调用，但是我用 VS2015（以及 opencv 3.1）测试发现，还是没法多线程同时调用啊，必须要加锁。 另外，您提供的 dll 似乎是 VS2010 编译的，而新版的接口里面用到了 vector，编译器版本不匹配时在 vector 析构的时候就会出现问题，除非调用之前为 vector 预留足够的空间。

我在facedetectcnn-model.cpp中找到了facedetect_cnn函数的定义，注释写着“input image, it must be RGB (three-channel) image!”而在头文件中facedetectcnn.h写着"input image, it must be BGR (three channels) insteed of RGB image!"。我觉得虽然是个小问题，但不一致的代码注释可能引起阅读混乱？

Description: Recently, we updated the commit of bitcoin for MSVC RWC testing, it failed to build due to the error C2039 and error C3861 like below, could you please help take a look? Thanks.

F:\gitP\ShiqiYu\libfacedetection\src\facedetectcnn.cpp(638,17): error C2039: 'to_string': is not a member of 'std' [F:\gitP\ShiqiYu\libfacedetection\build_amd64\facedetection.vcxproj]
         F:\gitP\ShiqiYu\libfacedetection\src\facedetectcnn.cpp(638,26): error C3861: 'to_string': identifier not found [F:\gitP\ShiqiYu\libfacedetection\build_amd64\facedetection.vcxproj]

Expected behavior: Build successfully.

Actual behavior: Build failed with error C2039 and error C3861

To reprodece:

1. open VS2019 x64 Native Tools command.
2. git clone https://github.com/ShiqiYu/libfacedetection F:\gitP\ShiqiYu\libfacedetection
3. mkdir F:\gitP\ShiqiYu\libfacedetection\build_amd64 and cd F:\gitP\ShiqiYu\libfacedetection\build_amd64
4. cmake -G "Visual Studio 16 2019" -A x64 -DCMAKE_SYSTEM_VERSION=10.0.18362.0 -DCMAKE_INSTALL_PREFIX=install -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DDEMO=OFF ..
5. msbuild /m /p:Platform=x64 /p:Configuration=Release libfacedetection.sln /t:Rebuild

More info: The commit of Libfacedetection we use: 4092665 VS version: VS2019 (v16.11.20)

Detailed log: build.log

I ran the detect-camera.cpp in example, and the average time consumption of face detection is 150ms. The environment is:

• AMD Ryzen 5 3550H
• 16GB RAM
• Microsoft Visual Studio 2022

By the way, I have enabled AVX2 and -O2 optimization when compling.

教授您好，有鉴于您的测试性能与精度相当地高，我使用您的模型测试FDDB的结果如下(红框部分)：

我想请问教授这个测试结果是正常的吗，我有看到有些评测说您的模型可达到 0.994 平均精度(AP)，但我的测试都只达到0.968左右，

我目前是个程式菜鸟，有些地方不太了解其中的意义，如果有什么地方有做错或少做的话请您给与指教，谢谢！

Good morning. How can I implement processing of upside-down images? Looks like keypoints are the same, so I can't just check them and rotate image. Is there any tool or hint to work with that images?

Thank you.