Unsupervised Monocular Depth Estimation from Light Field Image.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Supervised learning dominates light field depth estimation, requiring extensive ground-truth data.
• Acquiring accurate depth maps for light fields is challenging, often limited to synthetic datasets.

Purpose of the Study:

• To develop an unsupervised monocular depth estimation network for light fields.
• To eliminate the dependency on ground-truth depth information for training.

Main Methods:

• Exploiting multi-orientation epipolar geometry inherent in light fields.
• Proposing a novel unsupervised network that predicts depth from the central view.
• Introducing three unsupervised loss functions: photometric, defocus, and symmetry loss.

Main Results:

• Achieved satisfactory performance on a public 4D light field synthetic dataset.
• Demonstrated superiority compared to state-of-the-art unsupervised methods.
• Validated effectiveness and generality on real-world light-field images.

Conclusions:

• The proposed unsupervised method effectively estimates depth from light fields without ground-truth data.
• The novel loss functions contribute to robust and generalizable depth estimation.
• This work represents a significant advancement in unsupervised light field depth estimation.