EPI Light Field Depth Estimation Based on a Directional Relationship Model and Multiviewpoint Attention Mechanism

Area of Science:

• Computer Vision
• Image Processing

Background:

• Light field (LF) image depth estimation is crucial for 3D reconstruction and target tracking.
• Challenges include occlusions, edge ambiguities, and noise, despite LF refocusing properties.
• Epipolar plane images (EPIs) offer a robust approach due to multidirectionality and pixel consistency.

Purpose of the Study:

• To develop a novel algorithm for accurate LF depth estimation using EPIs.
• To address limitations of existing subaperture-based methods.
• To improve robustness against noise and regional ambiguities.

Main Methods:

• Utilizes epipolar plane images (EPIs) as input.
• Employs a directional relationship model to extract directional features from horizontal and vertical EPIs.
• Integrates a multiviewpoint attention mechanism (channel and spatial attention) to emphasize EPI slope information.
• Applies residual modules with small stride convolutions to refine features and preserve critical EPI slope data.

Main Results:

• The proposed EPI-based algorithm demonstrates superior performance compared to existing methods.
• Achieved higher accuracy in depth estimation tasks.
• Effectively handles challenging regions like occlusions and edges.

Conclusions:

• The EPI-based depth estimation algorithm offers a significant advancement in LF image processing.
• The integration of directional models and attention mechanisms enhances accuracy and robustness.
• This method provides a promising solution for various LF applications requiring precise depth information.