Learning Structure-from-Motion with Graph Attention Networks