Embedding label structures for fine-grained features representation

Two contributions

1. jointly optimize the classification loss(softmax)  and the similarity loss(triplet) om CNN, which can generate both categorization results

and discriminative feature representations.

2.embed label structures(make, model and year of cars) or attributes(ingredients of food)

siamese network

combine softmax and contrastive loss in CNN via joint optimization

learn a face classify first, and then use  the triplet constraint to fine-tune and boost the performance(deep face recognition 2015 bmvc)

ｔｒｉｐｌｅｔ:类内差距小于类间差距就好，triplet can preserve the intra-class variation

Regarding the sampling strategy, methods in Facenet( A unified embedding for face recognition and clustering), or employ hard mining approaches to explore challenging examples in the training data

Testing stage: generates classification result through the softmax layer, or the fine-grained feature representation after the l2 normalization

Embed label structures

hierarchical labels