Disclamer: I don't know nothing about Neural Networks or deep learning.
VLAD and Fisher Vectors with Product Quantization are a good example of how Approximate Nearest Neighbor can be used for good image retrieval. However, it has been proved how Neural Netowrks approaches are more precise than Bag of Words or VLAD.
My questions are:
- Creating an NN descriptor is faster than classic approaches (such as VLAD or Fisher Vectors)?
- Can we use classic ANN approaches with NN codes for finding similar images? If not, how NN codes are indexed?
- Can we use metric distances for measuring similarity between NN codes?
