mutual loss about alignedreid-re-production-pytorch

[huanghoujing]

I have tried simple combination of classification loss and mutual loss, yet also in another project. I used two ResNet-50. The CMC Rank-1 accuray of baseline is ~84%; adding mutual loss gains over 1 point improvement.

In this project, GL + IDL + ML + TWGD has additional probability mutual loss and global distance mutual loss than baseline GL + IDL + TWGD. I did not try adding only probability mutual loss here.

[Phoebe-star]

hi
I'm having some troubles

1. in the paper, the classification loss is how to compute?
   input a image (224x224) to resnet 50 , the output is 7x7x2048 feature , and then? I have no idea
   is it using fully conv and softmax to calssification?

2. how about mutual loss? Is it just combine two different loss
   ( for example, triplet loss + classification loss = mutual loss

   how about metric loss?

   thanks you

[huanghoujing]

1. You can look into the structure of ResNet-50. For input of size 224x224x3, the output of ResNet-50 conv5 is 7x7x2048. Then there is a average pooling layer reducing it to 1x1x2048. Then there is the final FC layer. Besides, classification loss is calculated as in object recognition task.

2. Mutual loss is not the sum of losses. You can refer to paper Deep Mutual Learning.