[ECCV 2020] Training neural networks to predict visual overlap of images, through interpretable non-metric box embeddings