liuyuisanai / coco_loss Goto Github PK

In your arXiv version, it shows that this paper is accepted by NIPS Workshop 2017, while the current README of this project shows that this paper is accepted by NIPS.
NIPS and NIPS Workshop are quite different. The citation section in the current README is misleading and should be corrected?

Would it be any training code for COCO loss available soon?
Will you provide trained models for other tasks such as person recognition or face recognition?

Thanks

Thanks for the great idea of using classification layer for cosine distance computation!
Could you please elaborate on how you train that layer? As I understand you initialize weights with average feature vectors for each class and what do you do with biases? Do you use them for training?

Besides your 99.78% on lfw and 76.57% on megaface, I was also very curious about your baseline method. Training on casia, softmax can get 99.5% on lfw, only using resnet 101. As we know softmax generally get 97.+% for the other papers' baseline, can you share the detail of your network and training tricks to us?

your coco_loss,centerloss ;Can they be used for facial feature extraction?
Do I need to increase the number of network layers?

there are few questions I want to know:
1.
As I know there are 10 error pairs in all the 6000 pairs,so the upper limit should be 1-(10/6000)%=99.83%,and the accuracy in your paper is 99.86%
2.
Can you show you training dataset and network which make the softmax loss to get a extremly high accuary(99.75%) on LFW? Softmax loss is not a very strong constrained loss and the result should be 98+% while training by MS_celeb_1M.
3.
What's the difference between your COCO_Loss and the combination of ASoftmax_loss and L2-constrain_loss?

In your paper, it's said that the centroid of class as average of features over a mini-batch which implies you update the centroid during training according to different mini-batches. But it's also said that "the features are initialized from pretrained models and the initial value of centroid is thereby obtained". As the centroid is determined by a mini-batch, why a initial value is needed?

Hi authors, I just reviewed coco-loss-v2 paper and saw that you reached 99.78%.
What I want to know is did you remove the three duplicated entities between Webface dataset and LFW?

where is convert_img2caffe,i can't find it in your repo