seetafaceengine / seetaface2 Goto Github PK

请问下300W是68个人脸关键点，在计算定位误差时，你的81和5时怎么对应的？
五点坐标是直接合成眼睛和嘴巴关键点吗？81点坐标比68点多了哪些点，81点中的68点是不是和300w的一一对应的？

cmake --build .失败，报了2个错误：
“D:\SeetaFace2-master2\build\ALL_BUILD.vcxproj"(默认目标)（1） ->
“D:\SeetaFace2-master2\build\Facedetector\SeetaFaceDetector.vcxproj"(默认目标)（3） ->
“D:\SeetaFace2-master2\build\SeetaNet\SeetaNet.vcxproj"(默认目标)（4） ->
(C1Compile 目标)->
D:\SeetaFace2-master2\SeetaNet\src\include_inner\SeetaNetSimd.h(775): error C2719: "value" : 要求对齐 [D:\Seet
aFace2-master2\build\SeetaNet\SeetaNet.vcxproj]
D:\SeetaFace2-master2\SeetaNet\src\include_inner\SeetaNetSimd.h(847): error C2719: "value" : 要求对齐 [D:\Seet
aFace2-master2\build\SeetaNet\SeetaNet.vcxproj]

这个是什么错呢？

'core_size' is invalid when init a FaceDetectorPrivate using function below

FaceDetectorPrivate::FaceDetectorPrivate( const char *model_path, const CoreSize &core_size )
    : FaceDetectorPrivate( model_path, CoreSize( -1, -1 ), SEETA_DEVICE_AUTO, 0 )
{
}

x86_64，Ubuntu16.04，运行example报错：非法指令（核心已转储）

能把相应的转换矩阵给出来么

@seetafaceengine Thanks for opening your great work of face structure. Could you provide tool to
transform my own caffemodel of fr to your support format?

D:\VS2015\SeetaFace2-master\build>cd bin
系统找不到指定的路径。

I follow the readme step by step, but there is no bin folder in the final build folder.

I hope someone can help, I am very grateful.

how to use GPU？？

我想把初始化之后的facelandmark模型参数传递给其他函数，请问我该怎么做？
例如：
seeta::FaceLandmarker FL(FL_model);
//将FL传递给下面这个函数，FL应该就是初始化之后的facelandmark模型参数
getFacelandmark(FL, frame);

谢谢

无法打开文件“SeetaFaceRecognizer2d.lib”

如题，在测试seetaface人脸检测时，发现检测框会出现明显抖动现象，请问一下有什么方法来改善吗？

平台是选择的arm，但是编译的时候出现了找不到immintrin.h这个错误，想问一下大神们应该怎么去解决这个问题。

请问seeta face输入resnet网络的图片尺寸统一缩放到多少大小?

I would like to thank the authors for their times and efforts devoted to this open source project. I was in trouble when I attempted to extract features from two images and calculate their similarity. Here is my code :

It ran normally on the first part to extract the features of "1.jpg", but threw a exception when ran to the statement "auto points2 = FL.mark(simage2, face2.pos);". Here is the exception message:

I cannot figure it out what's wrong with my code.

when run example.cpp

for( SeetaFaceInfo &face : faces )
{
cout<< face.pos.x << " " << face.pos.y << " " << face.pos.width << " " << face.pos.height << " " << face.score << "\n";
}

"face.score" is the confidence of detection?

but when test fddb:the score look like strange
2002/08/11/big/img_591
1
186 71 170 204 460.925
2002/08/26/big/img_265
2
288 60 105 126 43.0423
53 32 103 123 2.36952
@seetafaceengine

如题，有相应的转换工具没

检测图片并提取特征花费需要4s多
1.开发环境：cpu：ARMv7 Processor rev 5 (v7l) 8核 内存：1g
2.测试程序增加-mfpu=neon编译跟没有加，提取特征时间基本一样
3.seetaface库跟程序都采用release版本
你好，请问如何优化呢

Could you please give more details of FDDB test setting? E.g., single scale or multi scale? what is the value of min face and three thresholds? Thank you.

ndk version: android-ndk-r10d
error info :
In file included from /data/project/SeetaFace2/buildarmv7/SeetaFace2/SeetaNet/jni/../src/orz/mem/vat.cpp:8:0:
/data/project/SeetaFace2/buildarmv7/SeetaFace2/SeetaNet/jni/../src/orz/mem/../tools/ctxmgr_lite_support.h: In static member function 'static void* seeta::orz::__thread_local_lite_context::swap(seeta::orz::thread_local_lite_context::context)':
/data/project/SeetaFace2/buildarmv7/SeetaFace2/SeetaNet/jni/../src/orz/mem/../tools/ctxmgr_lite_support.h:25:28: internal compiler error: in var_defined_without_dynamic_init, at cp/decl2.c:2811
auto pre_ctx = m_ctx;
^
Please submit a full bug report,
with preprocessed source if appropriate.
See http://source.android.com/source/report-bugs.html for instructions.
make: *** [/data/project/SeetaFace2/buildarmv7/SeetaFace2/SeetaNet/obj/local/armeabi-v7a/objs/seetanet2//src/orz/mem/vat.o] Error 1

请问README中说识别自己的人脸，需要修改example的底库注册列表，这是指哪一个文件，或是如何打开修改？
我使用的是windows10，vs2019

测试者如果想在底库中成功识别出自己的人脸，需要在example.cpp的底库注册列表部分添加以自己名称命名的图片(名称 + .jpg)，

Linux error:
pot.h:19:36: error: ‘function’ in namespace ‘std’ does not name a template type

cmake version 3.10.2, gcc version 7.4.0 (Ubuntu 7.4.0-1ubuntu1~18.04.1)
c++11 ???

3288 armv7 Ubuntu16.04 无法编译

./seeta/FaceEngine.h:3:32: fatal error: seeta/FaceDetector.h: No such file or directory

如题

我是只需要FaceDetector和FaceLandMarker和SeetaNet，所以在vs2015中只引入了这三个文件的源文件，然后在附加包目录下添加所需要的头文件所在的文件，自己编写了一个test的源文件，就报出该错误。求解。感谢

when I make SeetaNet, I got this error:
CMake Error at CMakeLists.txt:23 (if):
if given arguments:
"STREQUAL" "x86"
Unknown arguments specified

Thanks

用人脸库实际数据测了一下，大部分人都认错了，人脸库才800多不到1000，还有女的认错另一个男的，差距太大，同样的图片百度完全没问题

在example/search目录下执行make报如下错：
./seeta/Struct_cv.h:3:33: fatal error: opencv2/core/core.hpp: No such file or directory
compilation terminated.
Makefile:22: recipe for target 'example.o' failed
make: *** [example.o] Error 1

Thanks for making source code open。
Is it evaluated on widerface？Since widerface is harder。
Thanks

index 0 is the maximum score? thanks very much.

检测部分比较快，初略看代码应该还是用的改良版的MTCNN，优化了网络结构和PNet的输入大小等。识别网络我看说明是用的ResNet50，有没有用类似于arcface之类的损失训练，或者做一些模型的改进？这部分比较耗时，初看没有insightface的r50跑的块（我实现的mxnet C++版本，开启MKL）。

将自己的图片存进注册列表时的具体步骤是什么呢，是不是我重复操作导致了这个问题，请问怎么清除注册列表的图片重新存入呢？

我在LFW上测试了一下seetaface2的精度，6000对数据的EER只有99.5左右，与当前的最佳水平有些差距，不知道放出来的模型是更加适用于**人还是怎样？

C:/Users/yoda/AppData/Local/Android/Sdk/ndk-bundle/build//../sources/cxx-stl/llvm-libc++/include\math.h:999:5: error: static_assert failed due to requirement '!(std::is_same<float, __result_type>::value &&
std::is_same<float, __result_type>::value)' ""
static_assert((!(std::is_same<_A1, __result_type>::value &&
^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
D:/sources/SeetaFace2/SeetaNet/sources/jni/../src/include_inner/layers\SeetaNetPowerCPU.h:127:28: note: in instantiation of function template specialization 'pow<float, float>' requested here
val = std::pow( val, m_power );
^
D:/sources/SeetaFace2/SeetaNet/sources/jni/../src/include_inner/layers\SeetaNetPowerCPU.h:10:5: note: in instantiation of member function 'SeetaNetPowerCPU::Process' requested here
SeetaNetPowerCPU();
^
D:/sources/SeetaFace2/SeetaNet/sources/jni/../src/include_inner/SeetaNetCreateLayerDetailCPU.h:193:24: note: in instantiation of member function 'SeetaNetPowerCPU::SeetaNetPowerCPU' requested here
output_layer = new SeetaNetPowerCPU();
^
D:/sources/SeetaFace2/SeetaNet/sources/jni/../src/include_inner\SeetaNetCreateLayerMapCPU.h:50:101: note: in instantiation of function template specialization 'CreatePowerFunctionCPU' requested here
FunctionMap.insert( std::pair<int32_t, CREATE_NET_PARSEFUNCTION >( seeta::Enum_PowerLayer, CreatePowerFunctionCPU ) );
^
D:/sources/SeetaFace2/SeetaNet/sources/jni/../src/include_inner\SeetaNetCreateLayerMapCPU.h:71:205: note: in instantiation of member function 'CreateLayerMapCPU::CreateFunctionMap' requested here
...std::map<int32_t, int( * )( SeetaNetBaseLayer &, SeetaNet_LayerParameter &, SeetaNetResource* )> CreateLayerMapCPU::m_parse_function_map = CreateLayerMapCPU::CreateFunctionMa...

使用的gcc为：arm-poky-linux-gnueabi-gcc 4.9.1 报错如下：
/SeetaFace2-master/SeetaNet/src/orz/sync/../tools/ctxmgr_lite_support.h: In static member function 'static void* seeta::orz::__thread_local_lite_context::swap(seeta::orz::__thread_local_lite_context::context)':
/SeetaFace2-master/SeetaNet/src/orz/sync/../tools/ctxmgr_lite_support.h:25:28: internal compiler error: in var_defined_without_dynamic_init, at cp/decl2.c:3005
auto pre_ctx = m_ctx;
^
Please submit a full bug report,
with preprocessed source if appropriate.

Multiple people appear in video. Can only one person be identified?

in ubuntu 18.04, compile not work. tried many methods. can you give a detailed steps about compiling?

如果有作持久化，持久化的文件在哪里，另外只看到注册的代码，没有看到有删除的

复现步骤：安卓平台，照相机采集数据输入，图片大小 640x360，NV21格式转成RGB24格式，可以识别出人脸；
// byteArray 是 NV21格式图片数据
jbyte* imgData = env->GetByteArrayElements(byteArray, 0);

SeetaImageData imageData;
imageData.width = width;
imageData.height = height;
imageData.channels = 3;
imageData.data = (unsigned char*)NV21ToBGR888(width, height, (const uint8_t*)imgData);

SeetaFaceInfoArray infoArray = faceDetector->detect(imageData);

if (infoArray.size == 0) {
    LOGI("No face found");
    env->ReleaseByteArrayElements(byteArray, imgData, 0);
    free(imageData.data);
    return (jboolean)false;
}

LOGI("A face found”);

输出为：A face found

但如果直接使用灰度图输入，无法识别出人脸。
// byteArray 是 NV21格式图片数据
jbyte* imgData = env->GetByteArrayElements(byteArray, 0);

SeetaImageData imageData;
imageData.width = width;
imageData.height = height;
imageData.channels = 1;
// NV21 头部为灰度图数据
imageData.data = (unsigned char*)imgData;

SeetaFaceInfoArray infoArray = faceDetector->detect(imageData);

if (infoArray.size == 0) {
    LOGI("No face found");
    env->ReleaseByteArrayElements(byteArray, imgData, 0);
    return (jboolean)false;
}

LOGI("A face found”);

输出为 No face found

在windows下使用cmake+cl编译时报如上错误

首先，我要说一句，这个项目的build 脚本 写的比较烂，可以说压根就没有在 Windows X64 平台测试过, 用Nmake 在windows平台编译肯定是失败的。

开源项目在开源前一定要做好跨平台的测试工作，否则发布出来了也是一堆乱七八糟的编译错误，让使用者对你这个项目的第一印象非常差 !!!

要想使用VS 2015 成功编译此代码， 你需要按照以下步骤做:

1. 安装一个MS VS2015 native build tools, 当然安装完整版本的VS 2015更好;
2. 安装 CMake (至少是3以上版本,我用的3.1.4)
3. 首先编译SeetaNet，然后编译FaceDector, FaceLandMarker, FaceRecognizer.
   编译的步骤都是一样的，先执行CMake 生成VC的工程文件，然后执行msbuild 生成dll和lib

3.1 cmake .. -G"Visual Studio 14 2015 win64"
这里的Visual Studio 14 代表 VS 2015, Visual Studio 15代表VS 2017, 以此类推。
win64是必须输入的，代表platform,不输入的话缺省值是win32, 在64位操作系统 执行build时
会出错。

3.2
msbuild ALL_BUILD.vcxproj /p:configuration=Release --生成release版本的dll
msbuild ALL_BUILD.vcxproj /p:configuration=Debug --生成debug版本的dll

4. 如果你在源码编译过程中还遇到了其它错误, 只能具体问题具体分析, 一般来说
   做到以下两步 基本就没问题了
   4.1 修改所有的vcxproj, sln文件，将Win32改成x64, -- 如果你的操作系统是64位的
   4.2 将所有的machine:X86 改成 machine:x64 -- 如果你的操作系统是64位的

In SeetaFace2/example/search/example.cpp, line 16, I find the setting:
seeta::ModelSetting::Device device = seeta::ModelSetting::CPU;
but when I change "CPU" to "GPU", nothing happened.
I have correctly install CUDA & CuDNN and used other DL frame like PyTorch.

请提供正确运行示例的操作系统版本