Cost Volume Pyramid Based Depth Inference for Multi-View Stereo.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Reconstruction

Background:

• Depth inference from multi-view images is crucial for 3D reconstruction.
• Existing methods often require high computational resources and fixed-resolution cost volumes.

Purpose of the Study:

• To develop a compact and efficient neural network for high-resolution depth inference.
• To improve depth map reconstruction results using a novel cost volume pyramid approach.
• To adapt the framework for unsupervised depth inference.

Main Methods:

• Proposed a cost volume-based neural network utilizing a cost volume pyramid built in a coarse-to-fine manner.
• Employed uniform sampling for initial cost volume construction and iterative refinement using depth estimates.
• Derived analytical geometric principles for compact cost volume pyramid construction.

Main Results:

• The cost volume pyramid approach results in a more compact and lightweight network.
• Achieved 6x faster inference speeds compared to state-of-the-art supervised methods with similar performance.
• Demonstrated superior performance in unsupervised depth inference scenarios.

Conclusions:

• The proposed cost volume pyramid network offers a significant improvement in efficiency and performance for depth inference.
• The framework is effective for both supervised and unsupervised depth estimation tasks.
• Enables high-resolution depth map generation with reduced computational cost.