Empirical Bayesian Light-Field Stereo Matching by Robust Pseudo Random Field Modeling.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Light-field stereo matching commonly uses Markov Random Fields (MRFs) for depth map inference.
• Previous MRF approaches often used fixed parameters, limiting adaptability to diverse image statistics and sampling densities.
• Explicit vision cues like depth consistency and occlusion were used for local adaptability, but confined applicability.

Purpose of the Study:

• To develop a novel empirical Bayesian framework for robust and broadly applicable light-field stereo matching.
• To address the limitations of fixed MRF parameters and confined applicability of previous methods.
• To enable depth estimation adaptable to dense, sparse, and denoised light fields, using intrinsic statistical cues.

Main Methods:

• Developed a Robust Pseudo Random Field (RPF) framework based on pseudo-likelihoods and hidden soft-decision priors.
• Employed soft expectation-maximization (EM) for model fitting and hard EM for robust depth estimation.
• Introduced novel pixel difference models for simultaneous adaptability and robustness.

Main Results:

• The proposed stereo matching algorithm demonstrates robust estimation of scene-dependent parameters and quick convergence.
• Achieved superior depth accuracy and computation speed compared to state-of-the-art algorithms.
• The framework is applicable to dense, sparse, and denoised light fields, and both true-color and grey-scale pixels.

Conclusions:

• The Robust Pseudo Random Field framework offers a significant advancement in light-field stereo matching.
• The method provides broad applicability and robust performance across various light-field conditions.
• Outperforms existing techniques in both accuracy and efficiency for depth map estimation.