layumi / 2015_face_detection

CVPR2015 Cascade CNNs for Face Detection

Hi there 👋

Contact Me:

✉ Email: zhedongzheng AT um.edu.mo

✧ Website: http://zdzheng.xyz

✧ Linkedin: https://www.linkedin.com/in/zhedongzheng/

✧ Google Scholar: https://scholar.google.com/citations?hl=en&user=XT17oUEAAAAJ

2015_face_detection's Issues

The version of matconvnet

What's the version you are using for this project? I tried the newest version but it failed with the error:

Error in vl_simplenn (line 182)
res(i+1).x = vl_nnconv(res(i).x, l.weights{1}, l.weights{2}, 'pad', l.pad, 'stride', l.stride) ;

Thanks!

No way to test Demo , ask why ?

No way to test Demo , ask why ?
error follow as down：

Thanks~

Error using distcomp.remoteparfor/getCompleteIntervals (line 138)
An unexpected error occurred during PARFOR: Error in remote execution of remoteParallelFunction : RUNTIME_ERROR

Error in vl_compilenn (line 405)
parfor i = 1:numel(mex_src)

Error in setup (line 15)
vl_compilenn('enableGpu', opts.useGpu, 'verbose', opts.verbose) ;

The client lost connection to worker 1. This might be due to network problems, or the interactive communicating job might have errored.

training process could not convergence

I test your training code on the AFLW dataset, but the 12-net could not convergence. Would I make some pre-processing on the train data? Thanks

demo run default

Does this code utilize Cudnn?

Hi,
Thanks for this amazing code.
I'm trying to detect faces in some images. The results are surprisingly accurate. I'm just wondering if I can improve its speed on my machine.
I've compiled matconvnet with gpu without Cudnn (My graphics card doesn't support Cudnn). Now the code takes ~5 seconds on each test image and there's a heavy load on my CPU while running the code. There's almost no difference in the performance when I compile matconvnet with CPU.
I don't know if the reason for this lack of speed is that I'm not using Cudnn (so I can buy a newer card with Cudnn support) or is it because of the computational load on CPU.

How to add 6-net??

how can i add new network like 6-net ?or 96-net

about the data of aflw

Hi,
I am reading your code of 'cascadeface_12net.m'. Would you please tell me how to prepare the data 'imdb12_v2_extended.mat' in detail? Thanks!

Best wishes,

Question about the implementation of FCN(heatmap)

Hi, layumi,
I am interested in the FCN method you adopt in 12-net as you mentioned in the technical report, but I get a little confused about the implementation. Could you please explain the point of deleting first col(row) of resized image in 12-net in scanpic_aflw_fast_slide.m? And why there isn't a subsequent 12-net-c?

Thanks!

how to train

thank you for sharing your codei am wondering how to complete the whole training process is there a demo or a document that can tell how to prepare the training data and train the model?
when training the model i need to separate the matlab matlab_12, utils and utils_12? and use them seperatively? what is the difference between the matlab and matlab_12?

net24 and net48 are gpuArray

Hi,
Thank you for your excellent work, clear code and easy to use and adapt.

On Test branch, pretrained models f24net-cpu.mat and f48net-cpu.mat are saved in gpu mode.
To avoid warning in test time on cpu, need to run:
NET = VL_SIMPLENN_MOVE(NET, 'cpu')

Custom and custom48 layer

Hi
Thank you for the code.
I have two questions here, one is about custom and custom48 layers. Why do you define custom and custom48? they seem just convolutional layers to me.
The second is about running on GPU. The running speed(3-4 seconds on large pic) is in GPU mode right? As long as the matconvnet is compiled in GPU mode, your code will run in GPU mode?
Thank you for the time!

请教一些问题

你好，我最近想用python实现一下这篇论文，但我的训练结果很差，有些问题想请教一下，如果可以的话希望留个邮箱~

Some questions about implementation and training

Hello. for how i work on rapid object detection problem. And I decided to use 2015 CVPR Cascade CNNs for Face Detection as base solution. I am developing it with theano\lasagne\python.
So now I have some questions.

• It is seems , that you use different architectures for networks( more conv layers, and one drop layer before output. Is it heuristics approach?
• How do you train detection nets? More concretely, witch patches do you use as negative examples? and how many examples do you use?
• Similar question about calibration net. How to choose distribution of 45 calibration classes?
• How do you implemented 24 net multi output? (see picture)
  
  If you want you can write answers for me by email [email protected] or just here
  Thank you and best wishes !

FDDB test cannot be repeated

Hi,
I have tried to rerun your work on FDDB, but I am not able to repeat your result. I based on your report to

1. in lines 49 (scanpic_fast_only12_24_48_newmodel_submean_demo),
   k=1:16
2. in lines 52 (scanpic_fast_only12_24_48_newmodel_submean_demo),
   ttt =ttt*1.18;
3. increase demo.m
   pad=20;
   However, I run out an error: res24 = vl_simplenn(net24, im24) ; when it detects on images, for example '2002/08/06/big/img_2526'
   I did tried to decrease the thres12 from .2 to .16 in lines 17 in scanpic_fast_only12_24_48_newmodel_submean_demo.m to avoid this error, As a result, the ROC curves is worse than Head and Hunter.
   Would you like to tell me how to repeat your experiment?
   Thanks
   Ho

vl_simplenn failure protection

vl_simple cannot handle empty input images, this is why I find this code useful:

if isempty(posbatch24.im)
   %disp('no24face');
   continue;
end
res24c = vl_simplenn(net24_c, posbatch24.im) ;

Do you think a similar protection can be useful in the following cases?

    res24 = vl_simplenn(net24, im24) ;
    res48 = vl_simplenn(net48, im48) ;
    res48c = vl_simplenn(net48_c, cim48) ;

Or in those cases the algorithm should never produce an empty input image?

Could you provide your training errors on each network?

Hi, thanks for your nice code. Recently I'm trying to implement the algorithm using Caffe. Could you provide your training errors on 12/24/48 net/calibration net each so I can check whether my code is correct?

Unable to run the code

The "Readme" file talks about some picture folder folder but I am unable to find any. Could you please help.

关于运行demo的问题

您好！是直接运行demo.m吗？我运行后没有什么反应，是需要在命令行输入mex那行命令吗？我是mac os 系统，谢谢您！
运行到了这一步：

In demo (line 19)
Warning: While loading an object of class 'gpuArray':
There is a problem with the CUDA driver or with this GPU device. Be sure that you have a supported GPU and that
the latest driver is installed.
In demo (line 19)