This paper proposes HF-Net, a hierarchical localization approach that uses a monolithic CNN to simultaneously predict local features and global descriptors for accurate 6-DoF localization. The method exploits a coarse-to-fine localization paradigm, first performing a global retrieval to obtain location hypotheses and then matching local features within those candidate places. This hierarchical approach significantly reduces runtime and enables real-time operation. By leveraging learned descriptors, the method achieves robust localization across large appearance variations and sets new state-of-the-art results on two challenging benchmarks for large-scale localization. The paper introduces a novel neural network, HF-Net, which jointly estimates local and global features, maximizing computation sharing. It is trained using multitask distillation, allowing the model to be flexible and efficient. HF-Net is shown to be more robust and accurate than existing methods, particularly in challenging conditions. The method is evaluated on three large-scale benchmarks, demonstrating its effectiveness in real-time applications. The results show that HF-Net outperforms existing methods in terms of accuracy and efficiency, particularly in challenging conditions. The paper also discusses the limitations of the approach, including the impact of model capacity on large, self-similar environments. Overall, the proposed method provides a robust and efficient solution for large-scale localization.This paper proposes HF-Net, a hierarchical localization approach that uses a monolithic CNN to simultaneously predict local features and global descriptors for accurate 6-DoF localization. The method exploits a coarse-to-fine localization paradigm, first performing a global retrieval to obtain location hypotheses and then matching local features within those candidate places. This hierarchical approach significantly reduces runtime and enables real-time operation. By leveraging learned descriptors, the method achieves robust localization across large appearance variations and sets new state-of-the-art results on two challenging benchmarks for large-scale localization. The paper introduces a novel neural network, HF-Net, which jointly estimates local and global features, maximizing computation sharing. It is trained using multitask distillation, allowing the model to be flexible and efficient. HF-Net is shown to be more robust and accurate than existing methods, particularly in challenging conditions. The method is evaluated on three large-scale benchmarks, demonstrating its effectiveness in real-time applications. The results show that HF-Net outperforms existing methods in terms of accuracy and efficiency, particularly in challenging conditions. The paper also discusses the limitations of the approach, including the impact of model capacity on large, self-similar environments. Overall, the proposed method provides a robust and efficient solution for large-scale localization.