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

Bug in face_preprocess_10kUS/create_negative_24c.py? #13

Closed hiyijian closed 8 years ago

hiyijian commented 8 years ago

line 60-62 after calling detect_face_12c_net, no sort by confidence, then call LocalNMS which require input bboxes are sorted by confidence

is that a bug?

And would you please to share ROC on benchmark?

anson0910 commented 8 years ago

Yes, I think that is a mistake! Sorry... my calibration models were trained improperly, and I currently haven't found time to train them correctly, but I have fixed the training .prototxt files.

Previous ROC results on the FDDB were about 10% lower than the original work, but I think fixing the calibration models will reduce the difference.

hiyijian commented 8 years ago

sorry, I didnt see any change in training .prototxt.

anson0910 commented 8 years ago

The train_val.prototxt files of the calibration nets in CNN_face_detection_models were changed to have a false mirror property in the data layer.

hiyijian commented 8 years ago

yes. I noticed that issuse you guys discussed previously. One more question please: line 560-563 in face_detection/face_detection_functions.py, both current_x and current_y time 2, why is that? As far as I am concerned, current_x and current_y are exactly the coord of top-left corner of the proposed window. there is no need to times 2?

anson0910 commented 8 years ago

The 2 corresponds to the pooling layer of the first 12 net, since the pooling layer scales the input image down by a factor of 2.

hiyijian commented 8 years ago

thank you