search

technicolor-research / subic

Tensorflow implementation of a supervised approach to learn highly compressed image representations
Other
27 stars 4 forks source link

Question about the method SUBIC #1

Open appleleaves opened 6 years ago

appleleaves commented 6 years ago

Thanks for publishing such an inspiring method and sharing the code. Using classification loss to deal with the dataset CIFAR10 with VGG, one should achieve a testing acc. more than 70%(In fact, it should be higher), which has only 10 output number in the final fc layer. In this way, if I set your block size to 10, and number of blocks to 1, can I achieve something like 70%?

-Leaf

himalayajain commented 6 years ago

Hi,

If you are talking about classification accuracy then yes, it should achieve something like 70% or whatever the performance is. But this accuracy is not same as category retrieval accuracy, where the ranking is very important.

-Himalaya