Real-Time Fast 3D Reconstruction of Heritage Buildings Based on 3D Gaussian Splashing
Weijia Wang
Abstract
With artificial intelligence and deep learning related technologies sweeping through major industrial fields, deep learning-based 3D reconstruction algorithms have gradually taken over the mainstream due to their efficient, fast and realistic modelling effects, which are of great significance as they provide brand new ways and means for the protection, research and inheritance of heritage buildings. Most of the reconstruction work of heritage buildings is based on the motion recovery structure-multi-view stereo vision algorithm, which requires a large number of continuous photographs and has high computing cost, and it is difficult to meet the rapid reconstruction and updating of large scenes with multiple details. To address this problem, this paper introduces a Gaussian splash method for fast 3D reconstruction, voxel construction and splash rendering by means of continuous photographs of near and far scenes collected by UAVs. This method can achieve high-precision real-time rendering reconstruction of cultural relics and buildings in large scenes in about 20 min, and at the same time, it avoids the emergence of holes, and the restoration of colours, structures, textures, etc. is more realistic compared with the traditional methods, which provides a new way of thinking for the 3D reconstruction of cultural relics and buildings.