# NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior

**Authors:** Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, Victor Adrian Prisacariu  
**arXiv:** [2212.07388](https://arxiv.org/abs/2212.07388)  
**Date:** 2022-12-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.07388v1/

## Key points

- Paper title: "NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior"
- Addresses the challenge of training a Neural Radiance Field (NeRF) without pre-computed camera poses
- Previous methods struggle with dramatic camera movements
- Proposes incorporating undistorted monocular depth priors into the training process
- Depth priors generated by correcting scale and shift parameters during training to constrain relative poses between consecutive frames
- Introduces novel loss functions to achieve effective constraint
- Method can handle challenging camera trajectories that were problematic for existing techniques
- Experiments conducted on real-world indoor and outdoor scenes
- Method outperforms existing methods in terms of novel view rendering quality and pose estimation accuracy
- Incorporating undistorted monocular depth priors significantly improves NeRF model performance with challenging camera movements
- Valuable contribution to the field by addressing an important limitation in NeRF training
- Promising results and potential for further advancements in improving NeRF models' robustness and accuracy in applications involving dynamic camera movements.

## Summary

The paper titled "NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior" addresses the challenging task of training a Neural Radiance Field (NeRF) without pre-computed camera poses. Recent advancements have demonstrated the possibility of jointly optimizing NeRF and camera poses in forward-facing scenes, however they still struggle with dramatic camera movements. To overcome this limitation, the authors propose incorporating undistorted monocular depth priors into the training process. These depth priors are generated by correcting scale and shift parameters during training to constrain relative poses between consecutive frames. The authors introduce novel loss functions to achieve this constraint effectively. By leveraging these priors and constraints, their method can handle challenging camera trajectories that were previously problematic for existing techniques. To evaluate their approach, experiments are conducted on real-world indoor and outdoor scenes which demonstrate that their method outperforms existing methods in terms of both novel view rendering quality and pose estimation accuracy. This indicates that incorporating undistorted monocular depth priors can significantly improve the performance of NeRF models when dealing with challenging camera movements. Overall, this paper presents a valuable contribution to the field by addressing an important limitation in NeRF training. The proposed method shows promising results and opens up possibilities for further advancements in improving the robustness and accuracy of NeRF models in various applications involving dynamic camera movements.