about the data of aflw

[jinmeng]

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,

[layumi]

You can get face data from http://lrs.icg.tugraz.at/research/aflw/ following the instruction on website.
In this database, it provide bounding boxes for each face in the sqlite, which you can easily find and resize it 12x12 or 24x24 or 48x48 as positive sample. And negative samples can be extracted from any background picture (maybe some scene images on Imagenet).
And imdb.mat include data(pos&neg) and labels. The structure are as follow:
imdb -----> imdb.images -------> imdb.images.data (mat) like 12x12x3x25000(just put all pic together)
-------> imdb.images.labels (vector) like 1x25000
For example, you can show images by typing 'imshow(imdb.images.data(:,:,:,1));%the first face'
I set labels for neg is 2 and for pos is 1 which you can set it as you want.

[lgyhero]

@layumi Could you please show me your e-mail address? There are some problems I need ask you for help. Thanks!

[layumi]

My email is [email protected]

[luoyetx]

Hi @layumi , How to request this aflw dataset? Can a student request this dataset or must be a researcher or teacher can request this dataset? thx.

[layumi]

@luoyetx
I forget whether only professor can apply this dataset.
You may refer to http://lrs.icg.tugraz.at/research/aflw/

[luoyetx]

@layumi thx for your quick reply

[luoyetx]

@layumi I have sent a request for the dataset and got a quick reply with given user name and password. It can be easily accessed with .edu email.

[luanps]

Hi, i'm trying to train your 'cascadeface_12net', but i've had some problems (like a filters more larger than data). I use aflw faces and negatives samples (all cropped and resized to 12x12) and have a vector for labels (1 for pos and 2 for neg). Maybe my imdb has wrong, so could you please can tell me where is the bug in my code?

This is my "prepare_imdb.m":

clc
clear
f=dir('*.jpg')
files={f.name}
for k=1:numel(files)
      f = imread(files{k});
      fd = im2single(f);
      images.data(:,:,:,k) = fd;
    if ~isempty(findstr(files{k},'neg'))
        images.labels(k) =  2
    else images.labels(k) = 1
    end
end

clear f fd files k;
imdb.images = images;
save('imdb.mat');

Thanks so much!

[layumi]

@luanps It seems ok.
You may change 12net structure code in "f12net.m" which I didn't update to the latest.
There is a error in it. You may change filter size according to the test model.