关键词: 隐式神经表面, 三维重建, 截断符号距离场, 动态占用网格, 光度一致性约束

Abstract: To address the imbalance between efficiency and accuracy in neural implicit surface reconstruction caused by dense sampling and volumetric rendering, an efficient reconstruction method is proposed. First, dynamic occupancy grids with density thresholds were used to select effective sampling points, reducing redundant computation and memory usage. Then, multi-resolution hash encoding was integrated with the Truncated Signed Distance Field (TSDF) to enhance gradient stability and noise resistance through truncated distance constraints and hash feature interpolation. Finally, photometric consistency was enforced across views using Normalized Cross-Correlation（NCC）. On the NeRF-Synthetic dataset, The proposed method reduced training time by 98.8% compared to Neural Radiance Fields (NeRF), with significantly better surface reconstruction quality than Instant-NGP. On the DTU dataset, the method achieved a Chamfering distance(CD) and Peak Signal to Noise Ratio(PSNR), outperforming Instant-NGP by 1.21 mm and 2.93 dB, respectively. Though slightly behind State Of The Arts (SOTA) Neuralangelo (with CD higher by 0.02 mm and PSNR lower by 2.05 dB) while reducing training time by 97.5% compared to Neuralangelo. The results demonstrate its effectiveness in achieving fast and accurate 3D reconstruction in complex scenes.

Key words: implicit neural surface, 3D reconstruction, Truncated Signed Distance Field (TSDF), dynamic occupancy grid, photometric consistency constraint