Flux diagram for both localization algorithms discussed in this paper, UNEK and Centroid-Net. Both algorithms receive as input gamma-ray sky patches with resolution (64, 64, 5), then the U-Net is applied in order to obtain semantic information about the localization of sources. Finally, each algorithm uses a particular clustering algorithm to recover the position of sources. Note that the iterative k-means algorithm uses only one channel as an input, which majority voting algorithm uses three layers (see text for details).