A Survey on Learning to Hash

Motivation: Storage efficiency and Retrieval efficiency.

Locality sensitive hashing and learning to hash.

Data independent(LSH) and data dependent(Learning to Hash).

 Learning to hash:

pairwise similarity preserving, multiwise similarity preserving, implicit similarity preserving, and quantization.

Future:

End to end learning strategy for real application, directly learning the hash codes from the object, e.g., image, instead of first learning the representations and then learning the hash codes from the representation.

To improve the recall, two methods are adopted

1 take more hashing tables. 2 one hashing table and retrieval more buckets. 

Retrieval:

1 find the nearest buckets, and use other hashing codes or original feature to make a re-rank.