PNeRFLoc: Visual Localization with Point-based Neural Radiance Fields

TL;DR

PNeRFLoc introduces a unified point-based neural radiance field framework for improved visual localization, combining feature matching and view synthesis to enhance accuracy and robustness in outdoor environments.

Contribution

The paper proposes a novel point-based NeRF framework that integrates pose estimation and refinement, with a feature adaptation module and efficient rendering techniques for better localization.

Findings

Outperforms existing methods on synthetic datasets with well-trained NeRFs.

Achieves comparable results to state-of-the-art on benchmark datasets.

Enhances robustness against illumination changes and dynamic objects.

Abstract

Due to the ability to synthesize high-quality novel views, Neural Radiance Fields (NeRF) have been recently exploited to improve visual localization in a known environment. However, the existing methods mostly utilize NeRFs for data augmentation to improve the regression model training, and the performance on novel viewpoints and appearances is still limited due to the lack of geometric constraints. In this paper, we propose a novel visual localization framework, \ie, PNeRFLoc, based on a unified point-based representation. On the one hand, PNeRFLoc supports the initial pose estimation by matching 2D and 3D feature points as traditional structure-based methods; on the other hand, it also enables pose refinement with novel view synthesis using rendering-based optimization. Specifically, we propose a novel feature adaption module to close the gaps between the features for visual localization and neural rendering. To improve the efficacy and efficiency of neural rendering-based optimization, we also develop an efficient rendering-based framework with a warping loss function. Furthermore, several robustness techniques are developed to handle illumination changes and dynamic objects for outdoor scenarios. Experiments demonstrate that PNeRFLoc performs the best on synthetic data when the NeRF model can be well learned and performs on par with the SOTA method on the visual localization benchmark datasets.