alphaqi / mtcnn-light Goto Github PK

1、mtcnn::mtcnn(int row, int col)
float minl = row>col?row:col;
可能会导致最后一个尺度的图像的宽或高小于12，进而导致后面this->conv3_matrix的height或width小于0，导致feature2MatrixInit函数里报错。
按你的意思是不是应该row<col?row:col;

2、maxPooling
if ((pbox->width - kernelSize) % stride == 0)
有时候会报错，if(maxNum<(ptemp+i+kernelRowpbox->width))这里指针越界，我发现此时是(pbox->width - kernelSize) % stride == 0但是(pbox->height - kernelSize) % stride != 0。
是否应该加上heigth的限制即，
if ((pbox->width - kernelSize) % stride == 0 && (pbox->height - kernelSize) % stride == 0)

谢谢你的开源！
我用原来的训练的模型，检测近红图像效果不理想。是不是应该在近红外上再训练下呢？

any plan to add MTCNNv2 LNet which is using image patch around each landmark outputted from stage 3 to make a precise regression.

能否告知一下编码的格式（估计是中文？），是utf8和gb2312都试过了，还是显示不正常（e.g. line117 in network.cpp）。

Hi, I found that your test code is clock()/10e3, I think this is a bug, maybe is clock()/1e3, so the time maybe is 260ms?

the things I changed were:

1. compiling opencv from source reduced the runtime considerably from 50ms to 10ms on my system. I used the master branch. OpenCL was giving me problems with the opencv version install from the repositories.
2. Adding an optimization level of -O3 to cmake further increased the speed from 10ms to 5ms.

This was ran under fedora 27 on a i7-2600k processor. I do have a GTX 980 ti installed but I don't think OpenCL used it. opencv did not have cuda support in either version.

I guess that is doable using JNI. But I am not an expert on that. Anyone can provide a resource or guide on how to do that? Thanks a lot.

Your code is:
clock_t start; start = clock(); find.findFace(image); imshow("result", image); imwrite("result.jpg",image); start = clock() -start; cout<<"time is "<<start/10e3<<endl;
Why you compute the time of "ms" with "start/10e3"? i. e. How to compute the time of "ms" with C++? I think it is right with "start/1e3".

looking forward for your reply.

best.

请问编译windows版本所需要的openblas二进制包哪有下载？网上指的链接sourceforge里面的是源码非二进制包

What

Since you use INTER_LINEAR for resizing images in mtcnn,
I'd be happy if I could hear why not INTER_AREA from you.

If you have no policy for this interpolation, that will also be the enough answer to me!

Thanks in advance!!

Today I accidentially change the :

export OPENBLAS_NUM_THREADS=4

to

export OPENBLAS_NUM_THREADS=1
Than the detection timing getting better ?? !!

from 70 ms to 20 ms and the cpu usage decreased to %30 from %99..

I really confused ..

any idea why ?

raspbian latest is the OS. Rpi 3B+

首先area的计算大多数没有+1
first ，the area calculate not plus 1;
当我使用视频接入时，误检测较多，错误出现在
when I connect video，fp more，wrong below
firstBbox_.clear();
firstOrderScore_.clear();
secondBbox_.clear();
secondBboxScore_.clear();
thirdBbox_.clear();
thirdBboxScore_.clear();
正确写法eg:（right style）：
//second stage
...
firstBbox_.clear();
firstOrderScore_.clear();
if(count<1) return faces;
...
//third stage
secondBbox_.clear();
secondBboxScore_.clear();
if(count<1) return faces;
//after third stage
thirdBbox_.clear();
thirdBboxScore_.clear();
还有
mtcnn.cpp
bbox.x1 = round((striderow+0.5)/scale);
bbox.y1 = round((stridecol+0.5)/scale);
bbox.x2 = round((striderow+cellsize)/scale);
bbox.y2 = round((stridecol+cellsize)/scale);
network.cpp
if((*it).x2>height-1)(*it).x2 = height - 1;
if((*it).y2>width-1)(*it).y2 = width - 1;
may be better?

minsize = 60 , I can run 720P， but minsize = 40 ,I cannot run 720P, thx

when run the 4.jpg, no face is detected and error saying "on entry to sgemm parameter number 10 had an illegal value".
what do i miss?

If l set the pad is 1 when initing the conv1_wb like this initConvAndFc(this->conv1_wb, 20, 3, 3, 1, 1)
there will be an error: free(): invalid pointer: 0x00007fab8d8297d8
And I found in function convolutionInit(), when calculating the size of outpBox padding size is not considered.
I wonder if i can use padding based on this framework and how to ?
thank you very much

只能跑mtcnn的推理机？

Hi,

After a while , around few thousands frame it gives below error. I couldnt figure it out .

Is there anybody faced same problem ?

thx

OpenCV Error: Assertion failed (0 <= roi.x && 0 <= roi.width && roi.x + roi.width <= m.cols && 0 <= roi.y && 0 <= roi.height && roi.y + roi.height <= m.rows) in Mat, file /opt/concourse/worker/volumes/live/d8bcd4d1-79b2-4aa5-797a-b95097f1118f/volume/opencv_1512680501887/work/modules/core/src/matrix.cpp, line 538
/opt/concourse/worker/volumes/live/d8bcd4d1-79b2-4aa5-797a-b95097f1118f/volume/opencv_1512680501887/work/modules/core/src/matrix.cpp:538: error: (-215) 0 <= roi.x && 0 <= roi.width && roi.x + roi.width <= m.cols && 0 <= roi.y && 0 <= roi.height && roi.y + roi.height <= m.rows in function Mat

Hi,

Same code :

`while(true){
         start = clock();
         cap>>image;
         cv::resize(image, image, cv::Size(), 1.0 * FACE_DOWNSAMPLE_RATIO, 1.0 * FACE_DOWNSAMPLE_RATIO);

         find.findFace(image);

         imshow("result", image);
         if( waitKey(1)>=0 ) break;
         start = clock() -start;
         cout<<"time is  "<<start/10e3<<endl;
     }`

gives segmentation fault at raspberrypi3 . at

find.face(image)

any clue ?

Im using FreeBSD 13.1
openCV 4.5

I was able to compile the project.
Clueless on how to run the compiled binaries...

I get this error:

$ ./main /dev/video4

[ WARN:[email protected]] global /usr/ports/graphics/opencv/work/opencv-4.6.0/modules/imgcodecs/src/loadsave.cpp (239) findDecoder imread_('4.jpg'): can't open/read file: check file path/integrity

Abort trap (core dumped)

Thanks.

@AlphaQi as I know, mtcnn in cpu is slow, how do you speed it up?

Could you offer the train code?

您好，请问您的.txt 参数文件中 数据是怎么组织的？比如PNet的第一层卷积，权重是331*10，输入单通道图像，代码中使用了矩阵相乘，且将输入特征平面转换成了行向量的形式，那么这里的权重是按什么顺序写到.txt文件中的？多谢！

The problem still exists.
First of all ,I want to test 1920*1080 video.
Then,when i set the "minsize" to 60(default),the application output "the minsize is too small,please change it",and,I set the value of "minsize" to 80, "the feature2MatrixInit failed!!"Segmentation fault
And,I change the "mtcnn find(image.rows, image.cols)" to "mtcnn find(640,480)" in pikaqiu.cpp. Finally the program work correct. but I do not ensure the accuracy! I do not understand it.
BTW, I think the the way of timer is not correct,I am in ubuntu. your test code is clock()/10e3,maybe is clock()/CLOCKS_PER_SEC
Thanks

Hi,

in the http://image13.m1905.cn/uploadfile/2013/0205/20130205114820972_watermark.jpg image

there is false positives. how we can play the threshold for this ?

thx

最近我在尝试脱离openblas实现这个项目，不过结果总是有问题，不知道哪里有偏差，可能是openblas矩阵相乘那里理解有问题

when i use this code to process 1080p image ,it has an error.640*480 iamge is normal.

你好！我想问一下txt文件是怎么写入的？比如我用的是tensorflow的参数文件，一个卷积层的结构是（kernel，kernel， inputmap，outputmap），这四个维度应该按什么顺序写入txt？看了前面说按照reshape不行啊

如题

请问这个报错是什么原因呢？

有没有发现一些hard sample检测不出来？

代码中为：
clock_t start; start = clock(); find.findFace(image); imshow("result", image); imwrite("result.jpg",image); start = clock() -start; cout<<"time is "<<start/10e3<<endl;
有错误

应该先将start转换为秒，再转换为ms，转换为s要除以CLOCKS_PER_SEC
实际上1s=1000ms，从s到ms的转换应该是乘以1000.
正确代码应该为：
cout<<"time is "<<(double)start/CLOCKS_PER_SEC*1000<<" ms "<<endl;
得到人脸探测时间为400ms， 也就是0.4s