.png){kind=icon}
.png){kind=icon}
Return
GloNeRF: Boosting NeRF capabilities and multi-view consistency in low-light environments
.png){kind=icon}
PRE
AI
DOI：10.1016/j.cag.2025.104209.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Neural Radiance Field (NeRF) significantly enhances the photorealism and detail richness of images by precisely rendering complex scenes using deep learning models, offering revolutionary improvements in novel view synthesis and three-dimensional scene modeling. However, when processing images captured in low-light conditions, the performance of NeRF can be significantly compromised, resulting in the loss of details and a decline in image quality. Although simply applying 2D low-light enhancement methods can improve image quality, this approach may lead to inconsistencies across multi-views, thereby introducing floating artifacts in the reconstructed neural radiance field. To address this issue, we propose a new framework. Initially, we enhance a series of low-light images using 2D low-light enhancement techniques. Subsequently, after volumetric rendering, we apply a bilateral grid approximation in the process of low-light image enhancement. Finally, we assign a bilateral grid to each training view to accommodate changes induced by low-light enhancement. During the Inference phase, we remove the bilateral grid, directly rendering novel views to ensure consistency across multi-views. Extensive experiments were conducted on three types of low- light datasets, and the results demonstrated satisfactory performance in both qualitative and quantitative evaluations.
Keywords:
Neural radiance fields
Low-light enhancement
Bilateral grid
Multi-view consistency
Journal
C
IF:
2.8
Papers:
82
Citations:
4.3K