This paper proposes the use of Neural Radiance Fields (NeRF) as a scene representation for visual localization. NeRF provides a compact and realistic representation of 3D scenes, enabling accurate geometry and appearance modeling. The authors explore the potential of NeRF's internal features for establishing precise 2D-3D matches, which are crucial for visual localization. They introduce NeRFMatch, an advanced 2D-3D matching function that leverages NeRF's internal knowledge learned through view synthesis. The proposed method is evaluated on standard localization benchmarks, achieving competitive results for localization performance on the Cambridge Landmarks dataset. The paper discusses various approaches to visual localization, including structure-based methods, image retrieval, and learned localization techniques. It highlights the limitations of existing methods and the potential of NeRF as a more compact and interpretable representation. The authors compare their method with existing approaches, showing that NeRF features are effective for 2D-3D matching. They also conduct ablation studies to evaluate the impact of different components of their method, including different 3D features and matching functions. The results show that the proposed method achieves competitive performance on both indoor and outdoor datasets. The authors also discuss the limitations of their approach, including the performance gap when applied to the indoor 7-Scenes dataset. They conclude that NeRF features are highly effective for 2D-3D matching and that further research is needed to improve the performance of NeRF-based localization in indoor environments. The paper provides detailed implementation details and supplementary information on the proposed method.This paper proposes the use of Neural Radiance Fields (NeRF) as a scene representation for visual localization. NeRF provides a compact and realistic representation of 3D scenes, enabling accurate geometry and appearance modeling. The authors explore the potential of NeRF's internal features for establishing precise 2D-3D matches, which are crucial for visual localization. They introduce NeRFMatch, an advanced 2D-3D matching function that leverages NeRF's internal knowledge learned through view synthesis. The proposed method is evaluated on standard localization benchmarks, achieving competitive results for localization performance on the Cambridge Landmarks dataset. The paper discusses various approaches to visual localization, including structure-based methods, image retrieval, and learned localization techniques. It highlights the limitations of existing methods and the potential of NeRF as a more compact and interpretable representation. The authors compare their method with existing approaches, showing that NeRF features are effective for 2D-3D matching. They also conduct ablation studies to evaluate the impact of different components of their method, including different 3D features and matching functions. The results show that the proposed method achieves competitive performance on both indoor and outdoor datasets. The authors also discuss the limitations of their approach, including the performance gap when applied to the indoor 7-Scenes dataset. They conclude that NeRF features are highly effective for 2D-3D matching and that further research is needed to improve the performance of NeRF-based localization in indoor environments. The paper provides detailed implementation details and supplementary information on the proposed method.