Fig. 4.

Illustration of the proposed method. Left: a deep convolutional neural network which takes image patches of 256$\times$256 pixels as inputs, denoted as $f_{\mathit{loc}}$. Right: the aggregation network, $f_{\mathit{agg}}$, that takes the concatenation of three types of maps as inputs: 1) location indicator map, $\mathbf{I}$, in gray, generated by downscaling the binary mask indicating the cropping window’s location, 2) saliency maps, ${\mathbf{S}}_m$ and ${\mathbf{S}}_b$, in red and green, generated by the context network, based on Globally Aware Multiple Instance Classifier [16], 3) embedding map, $\mathbf{E}$, in blue, formed by the representation $\mathbf{h}$, produced by $f_{\mathit{loc}}$