This document reports the use of Graph Attention Networks for classifying oversegmented images, as well as a general procedure for generating oversegmented versions of image-based datasets. The code and learnt models for/from the experiments are available on github. The experiments were ran from June 2019 until December 2019. We obtained better results than the baseline models that uses geometric distance-based attention by using instead self attention, in a more sparsely connected graph network.