Fast and Accurate Depth Estimation from Sparse Light Fields.

Area of Science:

• Computer Vision
• Photogrammetry
• Computer Graphics

Background:

• Dense depth reconstruction is crucial for 3D scene understanding.
• Processing sparse, wide-baseline light field data presents significant challenges due to occlusions and textureless regions.
• Existing methods often require substantial computation time and memory.

Purpose of the Study:

• To develop a fast and accurate dense depth reconstruction method for sparse, wide-baseline light field data.
• To reduce computational and memory requirements while preserving geometric accuracy.
• To enable efficient depth map generation for complex scenes.

Main Methods:

• Over-segmenting images into non-overlapping superpixels, modeled as planar patches.
• Estimating initial depth maps via plane-sweeping per view.
• Jointly refining depth maps using belief-propagation-like optimization in the superpixel domain.
• Employing a synchronous message update schedule on parallel graphics hardware for efficient processing.

Main Results:

• Achieved significant reduction in memory and computation requirements.
• Preserved image geometry and object contours effectively.
• Demonstrated globally consistent dense depth maps with few refinement iterations, even in challenging regions.
• Obtained depth reconstruction in approximately one second per full high-definition view.

Conclusions:

• The proposed superpixel-based method offers a highly efficient and accurate solution for dense depth reconstruction from sparse light field data.
• The approach effectively handles textureless regions and occlusions, outperforming existing methods in speed while maintaining comparable accuracy.
• This technique has the potential to significantly advance real-time 3D scene analysis and applications requiring dense depth information.