star-clouds / centerface Goto Github PK

您好，请问您的训练代码什么时候能够开源，对应论文发表时间呢？ 非常感谢

首先感谢你们贡献如此出色的idea！但是我在阅读论文过程中碰到了一些问题
在论文3.1部分，你们在P3，P4，P5层后加上了FPN，这三个层对应于原图分别下采样8、16、32倍，但是在论文3.2中，又说取下采样4倍的layer加检测头，这让我十分疑惑，是否论文有些错误？
期待您的回复

cv dnn版有个奇怪的现象，当检测过摆头角度很大的人脸后，再检测其他人脸照片时 landmark混乱，甚至跑到人脸范围以外，请问什么原因？prj-opencv-cpp/用的onnx模型

When excute ./demo ../../models/onnx your_image_path
will return
terminate called after throwing an instance of 'cv::Exception'
what(): OpenCV(4.1.1) ../modules/dnn/src/onnx/onnx_importer.cpp:57: error: (-210:Unsupported format or combination of formats) Failed to parse onnx model in function 'ONNXImporter'
Do you know how to solve it?
Thanks!

Great Work..
But can you tell how to handle image quality issue
The model is detecting blur faces too and providing poor result
i want to eliminate blur faces
the score of blur face detected is high too

Source code is BSD, MIT, Apache or GPL?
And model files are under the same license?

I have tested this model on my own 5000 images and the accuracy seems to be half of the claimed.

5 faces detected in the following is a disaster.
https://ibb.co/PCL0qNv

请问使用大模型 centerface的精度能match pyramidbox吗？速度怎么样？您有做过实验吗？

Processing 4k images, memory consumption is very large, is there a way to reduce it?

in decode() part,

https://github.com/Star-Clouds/CenterFace/blob/master/prj-python/centerface.py#L63

x1, y1 = max(0, (c1[i] + o1 + 0.5) * 4 - s1 / 2), max(0, (c0[i] + o0 + 0.5) * 4 - s0 / 2)

what's the 0.6 means? why not (center_point_x + offset_x)*scale - width/2 ?

s0, s1 = np.exp(scale0[c0[i], c1[i]]) * 4, np.exp(scale1[c0[i], c1[i]]) * 4
it means
scale0[c0[i], c1[i]] = bbox_width/ img_width_before_resize ?

faceboxes的推理速度有吗？

why the key points are not calculated as described in the paper?

the regression of the five facial landmarks adopts the target normalization method based on the center position
however the code is based on the left-top point.
facebox.landmarks[2j] = x1 + lm[(2j+1)*spacial_size+index] * s1

It can be seen from Netron that the input stage of the network is 10 × 3 × 32 × 32. Is it convenient for someone to tell how this structure is implemented? Or when the corresponding paper of the code will be published?

How are you handling the loss for occluded landmarks?
I was looking for this detail in the paper but I didn't find anything.

@ywlife Your released paper mentioned that RetainFace introduced five levels of face image quality and annotates five landmarks on faces.
I only found 5 landmarks ground-truth from RetainFace(InsightFace github resposity), but did not find quality ground-truth.
How could I got it?

when i install the ncnn and test prj-ncnn in ubuntu cmake && make it , i encountered the following problems:
fatal error: net.h: No such file or directory
compilation terminated.
CMakeFiles/demo.dir/build.make:62: recipe for target 'CMakeFiles/demo.dir/cpp/ncnn_centerface.cpp.o' failed
who can tell me why this error occured, thanks

I'm trying to reproduce the results of the paper, and this is what I have now. Is this your training data set?

Easy   Val AP: 0.9230461666813543
Medium Val AP: 0.9050905330904577
Hard   Val AP: 0.679394673204941

thanks.

1、你这个代码是基于哪篇论文实现？
2、你会公开训练代码吗？

我想看下论文和训练代码，更深入的了解你的人脸检测方法

您好，我在macos上，从摄像头读取图片测试，时间大约是420ms。这个速度正常吗。

Does this work follow the CenterNet? What is the relationship between the two? Is there a paper to describe it?

this method from which paper,thanks

仓库里的python代码是基于cv2.dnn的，我尝试着去用onnxruntime，但模型代码没有给，有些参数不知道设，如果作者有时间的话，可以放下tvm和onnxruntime的推理demo

thanks for your project. and how to train on my own data? thanks.

error: (-210:Unsupported format or combination of formats) Failed to parse onnx model in function 'ONNXImporter'.
Ubuntu18.04, opencv4.1.0.
ncnn model is running ok, but onnx model can not run, always failing. can you help me? somebody help me out?

您好，我看您的图片demo上有 facial landmark detection，但是在demo中并没有看到相应的部分，想问一下您这一部分是用什么模型做的？

Thanks for your excellent work! For SIO(Single Inference on the Original) evaluation schema, I get bad results on widerface validation set. Maybe it is due to the suboptimal thresholds (thres_score, thres_nms, etc). Could you please share your thresholds for widerface evaluation? Thanks pretty much!

Hi, you get the better results than retinaface on WiderFace-Val, you input the image of the WiderFace-val in origin size? We know that the performace will be bettter when the size of input is larger.

I have below error.

Do you have comparison with the mtcnn cpp speed ?

./demo ../../models/onnx ~/Downloads/selfi.jpg 
libc++abi.dylib: terminating with uncaught exception of type cv::Exception: OpenCV(3.4.8) /Users/xxx/Downloads/opencv-3.4.8/modules/dnn/src/onnx/onnx_importer.cpp:327: error: (-215:Assertion failed) model_proto.has_graph() in function 'populateNet'

Good work but it is taking 7 CPU cores on a live camera.
I would like to see the performance difference if this application was using GPU.

CenterFace指的是mobilev2+centernet吗？ CenterFace-small是在此基础上做优化？

hi，你好
我将ncnn改了下做视频和摄像头的测试，但是跑一会后，cpu占用率会突然飙到96或以上，
是不是哪里没释放，泄漏了啊。

good job!

https://github.com/Star-Clouds/centerface/blob/5c59451430105a37d6069d68cdeaf0690fba674a/prj-opencv-cpp/cpp/cv_dnn_centerface.cpp#L163

Hi,

I compiled opencv with cuda support and compiled demo cpp-opencv.

its only using cpu all cores not gpu ?

how I can use gpu ?

Good job！And the centerface.onnx contained face alignment??

CMake Error at CMakeLists.txt:23 (set_target_properties):
set_target_properties Can not find target to add properties to: make

Are there any plans to release the small model?

转换成阿里MNN模型，在树梅派4b上，4线程，输入3x480x640，推理消耗大概340ms+，不知道small模型有多大提升呢？

你好，你公布的模型的速度是readme中的速度嘛？？

Hi ,
Your paper looks exiting but there is no training and data-set available, can you please provide the same so that we can evaluate the paper independently.

Regards
Amit

Your demos are very impressive.
What scoreThresh and nmsThresh did you set to create these?
The default ones?

我看到速度方面的测试结果是在RTX2080TI上测试的，但是opencv的DNN模块好像并不支持GPU？

作者你好，我在ubuntu 16.04下测试prj-ncnn时 make 报错，ncnn已经安装编译成功，cmake 时顺利通过，我的环境如下：ubuntu 16.04 , python3.6, opencv 4.1, cmake3.9.1， gcc5.4 ，protobuf2.6.1,请问这个项目对环境有具体要求吗？

Which inference framework is used for the inference latency on https://github.com/Star-Clouds/CenterFace#inference-latency