(PMDH16) Convolutional Patch Representations for Image Retrieval - an Unsupervised Approach
Posted: 2016-09-26 , Modified: 2016-09-26
Tags: neural net, vision
Posted: 2016-09-26 , Modified: 2016-09-26
Tags: neural net, vision
We write things in the continuous limit. In actuality we have to discretize.
Define the kernel \(K:(\Om \to \R)\times (\Om\to \R) \to \R\) by \[\begin{align} K(M,M') = \sum_{z,z'\in \Om} e^{-\rc{2\be^2}\ve{z-z'}^2} \ka (P_z, P_{z'}')\\ \ka(P_z,P_{z'}') & = \ve{P_z}\ve{P_{z'}'} e^{-\rc{2\al^2} \ve{\wt P_z - \wt P_{z'}}^2} \end{align}\]where \(P_z =M_{z+[0,e)\times [0,e)\times [0,d)}\).
Background: see kernels.