Robust 2.5D Feature Matching in Light Fields via a Learnable Parameterized Depth-Degraded Projection.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Accurate 2D image feature matching is hindered by the loss of 3D scene information.
• Existing methods struggle with depth ambiguity in feature matching tasks.

Purpose of the Study:

• To develop a novel 2.5D feature for robust and accurate image feature matching in computer vision.
• To address the challenge of depth ambiguity in light field feature matching.

Main Methods:

• Introduced a 2.5D feature leveraging Fourier Disparity Layer (FDL) disparity as a depth proxy.
• Proposed a parameterized depth-degraded projection for geometric transformation between light fields.
• Developed a simple network to estimate fundamental matrix, translation vector, and depth compensation term.

Main Results:

• The 2.5D feature effectively reduces depth ambiguity without explicit depth estimation.
• Achieved accurate feature matching by minimizing reprojection errors.
• Outperformed existing 2D and light field feature matching algorithms on a public dataset.

Conclusions:

• The proposed 2.5D feature matching approach offers superior accuracy and robustness.
• Light field disparity provides a valuable cue for improving feature matching.
• The method demonstrates significant advancements in computer vision applications requiring precise feature localization.