Cascade light field disparity estimation network based on unsupervised deep learning.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Light field disparity estimation is crucial for various applications.
• Efficiently processing high-dimensional light field data remains a challenge.
• Existing supervised methods require ground truth disparity maps for training.

Purpose of the Study:

• To develop an unsupervised deep learning network for light field disparity estimation.
• To address the limitations of supervised methods by eliminating the need for ground truth data.
• To fully explore geometric characteristics of sub-aperture images for accurate disparity prediction.

Main Methods:

• Proposed a light field disparity estimation network with a cascade cost volume architecture.
• Implemented a coarse-to-fine prediction strategy leveraging sub-aperture image geometry.
• Designed a combined unsupervised loss function including occlusion-aware photometric and edge-aware smoothness losses.

Main Results:

• The unsupervised network achieved superior performance compared to existing unsupervised methods.
• The proposed method demonstrated better generalizability than supervised approaches.
• The combined loss function improved performance in challenging occlusion and textureless regions.

Conclusions:

• The developed unsupervised network effectively estimates light field disparity without ground truth data.
• The cascade cost volume architecture and combined loss function enhance accuracy and robustness.
• This approach offers a promising alternative for light field disparity estimation in real-world applications.