Differentiable Mapping Networks: Learning Structured Map Representations for Sparse Visual Localization