Light-field-depth-estimation network based on epipolar geometry and image segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Reconstruction

Background:

• Accurate depth estimation is crucial for various applications, including robotics and augmented reality.
• Light-field cameras capture spatial and angular information, enabling depth perception but posing unique challenges for depth estimation.

Purpose of the Study:

• To develop a fast and accurate depth estimation method for light-field images.
• To leverage epipolar geometry and image segmentation within a convolutional neural network framework.

Main Methods:

• A convolutional neural network (CNN) architecture is proposed, integrating epipolar geometry for initial disparity map estimation.
• Multi-orientation epipolar images serve as input, with convolutional blocks adapted to varying disparities.
• Image segmentation is employed to extract edge information from central sub-aperture images.

Main Results:

• The proposed method achieves high performance across multiple quality assessment metrics on the HCI 4D Light Field Benchmark.
• The approach demonstrates effectiveness in generating accurate depth maps for real-world light-field images.
• The integration of epipolar geometry and image segmentation leads to a fast and accurate depth estimation pipeline.

Conclusions:

• The developed CNN effectively combines epipolar geometry and image segmentation for robust light-field depth estimation.
• The method offers a significant advancement in speed and accuracy compared to existing techniques.
• This approach shows promise for real-world applications requiring precise 3D scene understanding from light-field data.