# NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections

**Authors:** Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, Daniel Duckworth  
**arXiv:** [2008.02268](https://arxiv.org/abs/2008.02268)  
**Date:** 2020-08-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2008.02268v1/

## Key points

- Paper presents a learning-based method for synthesizing novel views of outdoor scenes using unstructured photo collections
- Builds on the concept of neural radiance fields (NeRF) to model scene density and color
- NeRF struggles with real-world phenomena like variable illumination and occluders in uncontrolled images
- Authors propose extensions to NeRF, collectively referred to as NeRF-W, to address these limitations
- NeRF-W enables accurate reconstructions from unstructured internet photo collections
- System applied to internet photo collections of famous landmarks, resulting in photorealistic and spatially consistent reconstructions
- Represents a significant improvement over existing methods
- Innovative approach leveraging deep learning techniques for generating realistic scene representations from uncontrolled image collections
- Potential applications in computer vision and graphics.

## Summary

The paper titled "NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections" presents a learning-based method for synthesizing novel views of complex outdoor scenes using unstructured collections of in-the-wild photographs. The authors build on the concept of neural radiance fields (NeRF), which employs a multilayer perceptron to model the volumetric density and color of a scene. While NeRF performs well on static subjects captured under controlled settings, it struggles to capture real-world phenomena like variable illumination or transient occluders in uncontrolled images. To address these limitations, the authors propose several extensions to NeRF, collectively referred to as NeRF-W. These extensions enable accurate reconstructions from unstructured image collections sourced from the internet. The authors demonstrate the effectiveness of their system by applying it to internet photo collections of famous landmarks. The resulting reconstructions are photorealistic and spatially consistent, even when faced with unknown and confounding factors. This represents a significant improvement over existing state-of-the-art methods. Overall, this paper introduces an innovative approach that leverages deep learning techniques to generate realistic scene representations from uncontrolled image collections. By addressing challenges such as variable illumination and occlusion, NeRF-W demonstrates its potential for various applications in computer vision and graphics.