Beyond Photometric Consistency: Geometry-Based Occlusion-Aware Unsupervised Light Field Disparity Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Photogrammetry

Background:

• Unsupervised light field depth estimation faces challenges with occlusions and noise.
• Existing methods often rely on photometric consistency, which fails in occluded regions.

Purpose of the Study:

• To develop an occlusion-aware unsupervised framework for robust light field depth estimation.
• To improve accuracy and boundary preservation in noisy and occluded areas.

Main Methods:

• Designed a geometry-based light field occlusion modeling using forward warping and backward EPI-line tracing.
• Introduced two novel occlusion-aware unsupervised losses: occlusion-aware SSIM and statistics-based EPI loss.
• Developed a framework that moves beyond the simple photometric consistency assumption.

Main Results:

• The proposed method significantly improves light field depth estimation accuracy in occluded and noisy regions.
• The framework demonstrates superior preservation of occlusion boundaries compared to existing methods.
• Achieved noise- and occlusion-invariant representations of light field data.

Conclusions:

• The occlusion-aware unsupervised framework effectively addresses limitations in current light field depth estimation.
• This approach offers a more robust solution for real-world applications involving occlusions and noise.
• The novel loss functions and occlusion modeling contribute to enhanced performance.