OPAL: Occlusion Pattern Aware Loss for Unsupervised Light Field Disparity Estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Supervised learning excels in light field disparity estimation but struggles with real-world generalization due to lack of ground truth.
• Unsupervised methods offer potential for better real-world performance and generalization.

Purpose of the Study:

• To develop an unsupervised light field disparity estimation method with superior generalization to real-world data.
• To improve accuracy on synthetic datasets and reduce computational requirements.

Main Methods:

• Introduced the Occlusion Pattern Aware Loss (OPAL) to encode light field occlusion patterns for disparity calculation.
• Proposed an EPI transformer and a gradient-based refinement module for precise disparity estimation.
• Developed an efficient network architecture with reduced parameters.

Main Results:

• OPAL enables accurate and robust disparity estimation by effectively managing occlusions.
• The proposed method significantly outperforms state-of-the-art unsupervised methods in accuracy.
• Demonstrated superior generalization capabilities on real-world data compared to supervised methods.
• Achieved higher training and inference efficiency than existing learning-based approaches.

Conclusions:

• Unsupervised light field disparity estimation can surpass supervised methods in generalization and accuracy.
• OPAL is a key component for effective occlusion handling and parameter reduction.
• The developed method offers an efficient and accurate solution for light field disparity estimation.