Miper-MVS: Multi-scale iterative probability estimation with refinement for efficient multi-view stereo.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Learning-based multi-view stereo (MVS) methods excel at depth estimation but often suffer from low efficiency due to 3D convolutions and high computational costs.
• Existing multi-stage MVS approaches struggle to balance efficiency and generalization performance.

Purpose of the Study:

• To develop a highly efficient and generalizable multi-view stereo reconstruction method.
• To address the computational inefficiency of current learning-based MVS techniques.

Main Methods:

• Proposed a novel multi-scale iterative probability estimation with refinement (Miper-MVS) method.
• Introduced a high-precision probability estimator using dilated-LSTM for depth probability distribution encoding.
• Developed an efficient interactive multi-scale update module for integrating multi-scale information and improving parallelism.
• Implemented a Pi-error Refinement module to convert depth errors into grayscale error maps for edge refinement using high-frequency information.

Main Results:

• The proposed Miper-MVS method achieved the best generalization performance on the Tanks & Temples benchmarks among the most efficient methods.
• Demonstrated highly competitive performance on the DTU benchmark.
• The method balances efficiency (runtime and memory) with high accuracy.

Conclusions:

• Miper-MVS offers a significant advancement in efficient and accurate multi-view stereo reconstruction.
• The novel modules effectively improve depth estimation accuracy and edge refinement.
• The approach provides a strong alternative for real-world 3D scene reconstruction applications.