The document discusses a state-of-the-art model on the CARS196 dataset that uses proxy NCA loss to train embeddings for fine-grained vehicle classification. It trains multiple embeddings, where the dimension of each embedding equals the number of meta-classes and the number of embeddings is L. The model achieves higher performance than a 48-model ensemble by viewing samples within a meta-class as sharing a latent attribute and enforcing diversity of clusters across embeddings. It also questions whether attribute labels are truly needed to enhance feature learning.

1. ECCV2018， sota on CARS196, 9 pages

3. The number of embeddings: L
The dimension of each embedding: D
D equals the number of meta-classes
Proxy NCA loss for training each embedding

4. The dimension of each embedding: D The number of embeddings: L

5. The dimension of each embedding: D The number of embeddings: L
D L
Higher than our SSR with
48 models ensemble

6. Discussion1:
Do we really
need attribute
to enhance feature
learning?
Samples within a meta class can
be viewed as sharing a latent
attribute. So meta classes
corresponds to randomized
attributes

7. Discussion2:
In hidden layers, we may expect some clusters within the dataset.
A cluster may be viewed as a meta class.
employing meta class = enforcing diversity of clustering?
Discussion3:
Encoding the original one-hot label into a sequential label.
Using L-2 loss (or KLDiv loss, etc.) for learning the embedding brings about a similar
improvement?