DB-MVSNet: Unsupervised multi-view 3D reconstruction algorithm with two branches

Abstract

Supervised Multi-view 3D reconstruction algorithm based on ground truth data has great limitations in data acquisition cost and method promotion. Several unsupervised 3D reconstruction algorithms have even achieved similar results with supervised MVSNet series algorithms. We continue the work on the basis of previous research. Our DB-MVSNet includes two main tasks: first, the model introduces two parallel branches including depth estimation branch and semantic cluster branch. Secondly, we introduce normal depth in depth estimation branch, and adopt NNet network to realize the consistent propagation. The experimental results show that DB-MVSNet improves the overall performance of unsupervised 3D reconstruction algorithm.