Answers are grounded in this paper and cite their sources.
This paper presents a novel LiDAR view synthesis framework based on Neural Radiance Fields (NeRF). Unlike prior methods that primarily treat LiDAR as an auxiliary supervision signal and are restricted to image-based outputs, the proposed approach directly synthesizes high-quality point cloud views through several key innovations. Specifically, an enhanced scene encoding is designed to improve the continuity of features while preserving high-frequency details. Furthermore, an improved U-Net architecture is integrated to effectively capture the characteristics of LiDAR scans, adding external attention mechanisms to extract richer spatial information and local features. A new regularization scheme and geometric supervision strategy are proposed to efficiently guide model training. Experimental results on the KITTI-360 dataset demonstrate that the method significantly surpasses existing techniques for static scenes: the Chamfer distance of the synthesized point clouds is reduced by 8.25%, the root mean square error (RMSE) decreases by 17.15%, the structural similarity index (SSIM) increases by 10.22%, and the mean absolute error (MAE) of intensity drops by 25.25%.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: