bobrown / ssd-centerloss Goto Github PK

I am glad to see your centerloss codes, but I confuse how to use it, can your show me a full train.prototxt?
A lot of projects online is about classification, so I need a detection prototxt.
Another qusetion, how to determine the ‘center_features’?
thanks very much!

After training a SSD model with standard multiboxloss, I want to use your center loss for fine-turning, but I dont know how to set a proper center_loss_weight, and cannot find the default value, could you tell me your setting?
thanks a lot.

I want use the loss on https://github.com/zeusees/FaceBoxes/blob/master/examples/faceboxes/train.prototxt
I concat conv2_3/incep , conv3_2 , conv4_2 as "mbox_center_conf"
center_features: 23
but, the error is:
multibox_center_loss_layer.cpp:299] Check failed: conf_center_pred_.count() == bottom[4]->count() (16062464 vs. 2818048)

Hi.
I saw the README.md. But I still don't understand the param
multibox_center_loss_param {
center_features: 16 # center_features represents the length of features that is equal to the length of object centers in each default box.
}

What's the meaning of center_features. How did u get the value of center_features: 16? Did u set it Randomly?

Thanks for your work. I added your code to my SSD and trained on my own data, but I found some problems.

1. After 50000 iters, the mbox_loss decreased to 5~8, but mAP was always around 0.1 and did not increase. I wonder whether you have encountered such problem and know possible cause?

2. I changed train.prototxt according to the steps, but I did not change test.prototxt because it seems that detection_out layer should also be rewritten so that it could handle the input from its previous layer. So is it right not changing test.prototxt ? Or should I rewrite detection_out?

I'm glad to see your work with centerloss,have you got some better performance with center loss than ohem in ssd?

Thank you for your sharing！
My question is, the change about step (3) "Getting the center_features of each default box" is also in train.prototxt? The param "num_output" about VOC2007 equals (20+1)*3=63. what about 64?