Two-view semantic 3D reconstruction of ancient Chinese architecture via match propagation and component optimization
TL;DR
A fast two-view semantic 3D reconstruction method via joint match propagation and component optimization that efficiently produces reliable results, increasing match counts by an order of magnitude within seconds and improving component accuracy by approximately 5 percentage points.
Abstract
Structure parsing and 3D reconstruction are crucial for the digital preservation of ancient Chinese architecture (ACA). However, due to interfering factors such as tightly connected components and repetitive textures, traditional image-based semantic segmentation and 3D reconstruction methods often fail to obtain reliable results from limited images. To address this issue, this paper proposes a fast two-view semantic 3D reconstruction method via joint match propagation and component optimization. First, the proposed method constructs an ACA structure parsing model by enhancing the UNet++ framework with spatial-channel attention, pixel shuffle and a weighted loss function. Then, starting from initial components and seed matches, it iteratively performs match propagation and component optimization under geometric constraints and structural priors, generating dense matches with accurately refined components. Experimental results show that the proposed method efficiently produces reliable results, increasing match counts by an order of magnitude within seconds and improving component accuracy by approximately 5 percentage points.