A Light Multi-View Stereo Method with Patch-Uncertainty Awareness.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Multi-view stereo (MVS) methods reconstruct 3D models from multiple images.
• Existing MVS techniques often neglect coarse-stage features, limiting reconstruction accuracy.
• Fixed depth sampling ranges in MVS can hinder precise depth estimation.

Purpose of the Study:

• To develop a novel learning-based MVS method addressing limitations in feature extraction and depth sampling.
• To improve the accuracy and efficiency of 3D point cloud generation using multi-view imagery.
• To reduce the computational cost associated with MVS reconstruction.

Main Methods:

• Proposed a coarse-feature-enhanced feature pyramid network with attention mechanisms for improved feature extraction.
• Introduced a patch-uncertainty-based adaptive depth sampling strategy for refined depth estimation.
• Integrated edge features into iterative cost volume construction to enhance reconstruction fidelity.

Main Results:

• The proposed method demonstrates competitive 3D reconstruction quality on benchmark datasets (DTU, Tanks and Temples).
• Achieved lower GPU memory consumption compared to existing learning-based MVS approaches.
• Enhanced contextual features and adaptive sampling led to improved depth estimation accuracy.

Conclusions:

• The novel learning-based MVS method effectively improves 3D reconstruction accuracy by leveraging enhanced features and adaptive sampling.
• The approach offers a balance between high-quality reconstruction and computational efficiency.
• This work contributes to advancing the field of 3D scene reconstruction from images.