DRI-MVSNet: A depth residual inference network for multi-view stereo images.

Area of Science:

• Computer Vision
• 3D Reconstruction

Background:

• Accurate 3D reconstruction is crucial for scene geometry restoration.
• Current methods struggle with memory demands, leading to inaccuracies.
• High-accuracy 3D scene reconstruction remains a significant challenge.

Purpose of the Study:

• To propose a novel network, DRI-MVSNet, for highly accurate 3D image reconstruction.
• To address memory constraints and improve the accuracy and completeness of reconstructed point clouds.

Main Methods:

• A cascaded depth residual inference network (DRI-MVSNet) is proposed.
• Utilizes a cross-view similarity-based feature map fusion module for residual inference.
• Incorporates a combined module for channel and spatial information processing, attention mechanisms, and residual prediction with non-uniform depth sampling.

Main Results:

• DRI-MVSNet demonstrates competitive performance on the DTU and Tanks & Temples datasets.
• Achieves significantly superior accuracy and completeness in reconstructed point clouds compared to state-of-the-art benchmarks.
• The proposed network effectively handles memory demands for improved 3D reconstruction.

Conclusions:

• DRI-MVSNet offers a significant advancement in 3D image reconstruction.
• The network's novel modules enhance feature representation and depth map generation.
• It provides a robust solution for accurate and complete 3D scene reconstruction.