A Framework for Learning Depth From a Flexible Subset of Dense and Sparse Light Field Views

Area of Science:

• Computer Vision
• Photogrammetry
• Machine Learning

Background:

• Accurate depth estimation from light fields is crucial for 3D scene reconstruction and augmented reality.
• Existing methods often struggle with sparse sampling, occlusions, and require prior disparity range knowledge.

Purpose of the Study:

• To propose a robust learning-based depth estimation framework for light fields.
• To improve accuracy in occluded regions and along object contours.
• To eliminate the need for predefined disparity range information.

Main Methods:

• A three-step framework: initial depth estimation, fusion with occlusion handling, and refinement.
• Utilizes a flexible subset of input views for estimation.
• Employs two warping error measures for accurate fusion in challenging regions.

Main Results:

• The proposed method achieves superior performance compared to state-of-the-art light field depth estimation techniques.
• Demonstrates high accuracy even with sparsely sampled light fields.
• Effectively handles occluded areas and object boundaries.

Conclusions:

• The developed framework offers a significant advancement in light field depth estimation.
• Its flexibility and accuracy make it suitable for various applications.
• Outperforms deep neural network-based approaches and methods requiring cost volume computation.