Workflow from Scientific Research

Spatial sampler uses attention from low-res image to sample the top-$k$ regions from the (original) high-res input. $\mathbf{x} _{l}$ gives a global view, while $\mathbf{x} _{h,k}$ provides local views at important regions of the original image $\mathbf{x} $. The final global average pooling removes spatial dimension of the features, which are combined and fed to the three-head GRU classifier. The heads correspond to low-res features, high-res features, and their concatenation, and are used to encourage strong learning feature at each resolution.

DOI