3D wireframe model reconstruction of buildings from multi-view images using neural implicit fields

Abstract

The 3D wireframe model provides concise structural information for building reconstruction. Traditional geometry-based methods are prone to noise or missing data in 3D data. To address these issues, this paper introduces Edge-NeRF, a 3D wireframe reconstruction pipeline using neural implicit fields. By leveraging 2D multi-view images and their edge maps as supervision, it enables self-supervised extraction of 3D wireframes, thus eliminating the need for extensive training on large-scale ground-truth 3D wireframes. Edge-NeRF constructs neural radiance fields and neural edge fields to optimize scene appearance and edge structure simultaneously, and then the wireframe model is fitted from coarse to fine based on the extracted 3D edge points. Furthermore, a synthetic multi-view image dataset of buildings with 3D wireframe ground truth annotations is introduced. Experimental results demonstrate that Edge-NeRF outperforms other geometry-based methods in all evaluation metrics.