the usage of clusters in the training about tiny HOT 6 CLOSED

Good question. I recommend training all 25 templates on every scale. My experience is that training templates only on corresponding scales performs slightly worse.

from tiny.

Thanks peiyun. I just want to reproduce the result of your paper base on the py caffe. But now it seems that the result has something wrong. I think I have achieved most of ideas of your paper except more batch size and ohem. But the mAP in WiderFace validation set is only 6%. Too many false positives and bad performance for easy face. Could you give me some suggestions about that? Hope to see your training code as soon as possible. Thank you!

from tiny.

Hello peiyun, I found that if I reuse the parameter in resnet between ohem layer and hard mining layer subsequently , the AP value drop down about 10 percent t. But all the ohem paper suggest to reuse parameter .
Another problem is should I put the use_global_stats to be true or false of bn layer in training? It seems that how to implement the ohem in bn layer in resnet is very different with vgg16.

from tiny.

We did not implement the online hard negative mining in the same way as the OHEM paper did. Please refer to "Online hard mining and balanced sampling" in the appendix of our paper for our implementation.

from tiny.

Hello peiyun, thanks for your clarification. I have another question is how to " adding context" in resnet network? Firstly , should we use concat layer or Eltwise layer to adding context? Secondly , when I try to adding conv2, conv3 and conv5 , I found that the value of conv5's feature map in resnet is much larger than conv2 and conv3. If I adding these layers together, it seems that the featuremaps of conv2 and conv3 can not take any effect on the final result. Should I scale the value to match conv5? Thanks!

from tiny.

Here is a visualization of our network: http://ethereon.github.io/netscope/#/gist/8a0d5ef37da9dc4cd611d178404b3641. Hopefully it will address your questions.

from tiny.