What is common choice of the number of features to train codebook for product quantization? #2
Description
For PQ, we have to train encoder (give PQ code for a given original feature) first.
If we use too small number of examples for training encoder, the encoder is not well aligned with real dataset. It may give as poor quantization results.
If we use too many number of examples for training encoder, it represents well for real dataset. The downside is time for training the codebook.
Suppose we use 1B SIFT features. How many examples are common for training PQ codebook?
Is the same reasoning works as well for OPQ, LOPQ?