GeoNLF: Geometry guided Pose-Free Neural LiDAR Fields

TL;DR

The paper tackles robust novel view synthesis and multi-view registration for sparse, large-scale LiDAR sequences under pose-free settings. It introduces GeoNLF, a hybrid framework that alternates global bundle-adjusting neural LiDAR field optimization with a graph-based pure geometric optimizer, guided by a selective-reweighting strategy and 3D geometry constraints. Poses are updated in a pose-free manner using SE(3) representations with decoupled rotation and translation through the exponential map, while a Graph-based Robust Chamfer Distance provides geometry-driven global alignment. On NuScenes and KITTI-360, GeoNLF achieves state-of-the-art performance in both NVS and pose estimation, demonstrating robustness to outliers and improved geometric consistency in challenging outdoor, sparse data settings.

Abstract

Although recent efforts have extended Neural Radiance Fields (NeRF) into LiDAR point cloud synthesis, the majority of existing works exhibit a strong dependence on precomputed poses. However, point cloud registration methods struggle to achieve precise global pose estimation, whereas previous pose-free NeRFs overlook geometric consistency in global reconstruction. In light of this, we explore the geometric insights of point clouds, which provide explicit registration priors for reconstruction. Based on this, we propose Geometry guided Neural LiDAR Fields(GeoNLF), a hybrid framework performing alternately global neural reconstruction and pure geometric pose optimization. Furthermore, NeRFs tend to overfit individual frames and easily get stuck in local minima under sparse-view inputs. To tackle this issue, we develop a selective-reweighting strategy and introduce geometric constraints for robust optimization. Extensive experiments on NuScenes and KITTI-360 datasets demonstrate the superiority of GeoNLF in both novel view synthesis and multi-view registration of low-frequency large-scale point clouds.

Figures (10)

• Figure 1: Registration results. Pairwise algorithms such as GeoTrans qin2022geometric and ICP besl1992method suffer from error accumulation and local mismatches. Multi-view methods like SGHR wang2023robust and MICP choi2015robust still manifest outlier poses. Previous gradient-based approaches LiDARNeRF-HASH heo2023robust lack geometric consistency. Our method effectively avoids outlier frames and achieves superior registration accuracy.
• Figure 2: Overview of our proposed GeoNLF. We alternatively execute global optimization of bundle-adjusting neural LiDAR fields and graph-based pure geometric optimization. By integrating selective-reweighting strategy and explicit geometric constraints derived from point clouds, GeoNLF implements outlier-aware and geometry-aware mechanisms.
• Figure 3: Geo-optimizer and its impact on pose optimization. Comparison of (a) and (b) shows Geo-optimizer prevents incorrect pose optimization of NeRF.
• Figure 4: Impact of selective-reweighting training strategy on pose optimization. (a) Frames with outlier poses exhibit significantly higher losses. With selective-reweighting, outlier frames maintain a relatively higher loss without overfitting. (b) After several training iterations, the pre-trained outlier-aware NeRF can provide globally consistent geometric optimization for outlier frames.
• Figure 5: Qualitative comparison of NVS. We compared GeoNLF with other pose-free methods and GeoTrans-assisted NeRF. Especially, GeoTrans fails on Nuscenes due to the inaccurate poses.