Efficient neural RGB-D indoor scene reconstruction based on normal features

Abstract

Reconstructing large-scale indoor scenes from 2D images to 3D models presents substantial challenges, particularly in handling texture-less regions and extensive scene sizes with both accuracy and efficiency. This paper introduces a novel method for efficient and high-quality geometric reconstruction of indoor scenes using RGB-D images. Our approach integrates normal features as prior information into the RGB-D data and employs a truncated signed distance function (TSDF) to represent scene surfaces. Combined with multi-resolution hash encoding, the proposed method achieves both high reconstruction quality and computational efficiency. Specifically, we estimate normal vectors from RGB images as feature priors to guide surface fitting. To address the inaccuracies of normal estimation in regions with small objects or complex geometric details, we incorporate depth information to better constrain the surface fitting process. Additionally, multi-resolution hash encoding is used to stratify sampling points, enabling rapid feature lookups via hash functions. Experimental results demonstrate that the proposed method significantly outperforms existing approaches in terms of both reconstruction quality and computational efficiency.