Are there any requirements for the PCA_whitening_data? about crow

[pumpikano]

Yes, ideally the data used to fit whitening will have the same distribution as the test data.

[pumpikano]

If you are benchmarking, the test data really should be separate from the data you use for whitening (otherwise you are testing on the training data). In practice, it is reasonable to use your index data to fit the whitening, although you could risk overfitting if your test-time query images are from a different distribution (e.g. mobile phone query images but professionally-shot index images).

[xiaozhi2015]

@pumpikano Much thx!

[xiaozhi2015]

Well, can I generate a subset from my test data as the pca_whitening_data? @pumpikano

[xiaozhi2015]

OK, much thx for your answer!