FDDB test cannot be repeated

[chichan01]

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

[layumi]

@chichan01 As the original paper said, on fddb we should enlarge bounding box. I enlarge it by mutiple height with 1.4. (It's really important! ) My advice is to y1-(y2-y1)*0.4. (and notice that y<0 is OK)

[chichan01]

Are you talking about the ellipse factor? I did do it in which the factor is [1.4738,0.9581]; where 1.4738 is the main axis which is height you mentioned.

would you like to tell me what do you mean y1 and y2?

[layumi]

@chichan01 I suppose that bounding box can be represented as (x1,y1)left-top point (x2,y2) right-down point To enlarge the bounding box, you may change y1 = y1-(y2-y1)*0.4

[chichan01]

well, by doing that, the centre of the box will move up as well. Anyway, I will try.

[layumi]

@chichan01 I suggest that you may visualise the bounding boxes on image. It can help you to choose a better enlarge factor. I get 0.4 height by using this method.

[chichan01]

Would you like to confirm that the parameters of the networks you provide (in https://github.com/layumi/2015_Face_Detection/tree/test) are exactly you use to perform FDDB test? net12 = load('12net-newborn/f12net.mat') ; net12_c = load('12net-cc-v1/f12net_c.mat') ; net24 = load('24net-newborn/f24net-cpu.mat') ; net24_c = load('24net-cc-v1-no256/f24netc.mat') ; net48 = load('48net-6hard/f48net-cpu.mat') ; net48_c = load('48net-cc-cifar-v2-submean/f48netc.mat') ;

[layumi]

@chichan01 Yes, they're the latest model which I trained.

[chichan01]

Thanks for your response and I am able to reproduce the result but it seems that the true positive rate of this approach at less than 100 false positive is lower than that of head and hunt a lot.

[layumi]

@chichan01 Yes, this code didn't get the same result. As far as I concerned, some false-costive sample is selected. So the curve didn't perform well at first 50.

[chichan01]

What does false-costive sample mean? Besides, do you think to use big face dataset to improve your network? for example, you can combine your dataset with http://personal.ie.cuhk.edu.hk/~lz013/projects/CelebA.html (202,599 face images with landmark information)