Robust Attentional Aggregation of Deep Feature Sets for Multi-view 3D Reconstruction