search

anson0910 / CNN_face_detection

Implementation based on the paper Li et al., “A Convolutional Neural Network Cascade for Face Detection, ” 2015 CVPR
253 stars 147 forks source link

Expected result of running face_cascade_fullconv_single_crop_single_image.py script #22

Closed facenx closed 8 years ago

facenx commented 8 years ago

Hi,

I ran face_cascade_fullconv_single_crop_single_image.py script as noticed in readme without any modifications and got this result: image

Can you say if it is expected result? Or any setup should be done before running this script.

Thanks.

anson0910 commented 8 years ago

Yes, if the threshold is set lower to have higher recall rate, more rectangles should appear. However, as mentioned in this issue, the rectangles should first be sorted by confidence before NMS is performed, this is an error of mine.

facenx commented 8 years ago

Year. I found the absence of sorting at the end of _cal_face_24cnet function, but fixing doesn't appear something new: image I also played with threshold and found the best value 0.92 that leads to this result: image As before it seems not perfect especially because of simple example and hard threshold tuning on it. So do u see something like this on your screen, or I something do not yet know? :)

I do hope I am not putting you to too much trouble.

Thanks.

anson0910 commented 8 years ago

Yes, I did see something like this on my screen! No problem.

facenx commented 8 years ago

It seems that current quality is not enough to real usage. But code base has a big potential to create competitive detector with such production ready solutions as dlib, libccv etc. Thanks a lot.

anson0910 commented 8 years ago

No problem, and thanks for your comments!