UCGR: Closing the Discretization Gap in Light Field Depth Estimation via Unified Continuous Geometry Representation

Area of Science:

• Computer Vision
• 3D Geometry
• Machine Learning

Background:

• Light field (LF) cameras capture spatial-angular data for depth estimation, vital for 3D reconstruction, refocusing, and VR.
• Current deep learning methods struggle with the discretization gap between continuous scene geometry and discrete image sampling, impacting high-precision applications.
• This gap causes spatial ambiguities and depth inaccuracies, limiting LF depth estimation performance.

Purpose of the Study:

• To propose a novel representation and network for accurate and consistent light field depth estimation.
• To address the challenges posed by spatial and depth discretization in LF data.
• To enhance the effectiveness of LF depth estimation for demanding applications.

Main Methods:

• Introduced Unified Continuous Geometry Representation (UCGR) modeling scene geometry as a continuous field.
• Developed an Adaptive Plane Sampling Operator to preserve geometric details and mitigate spatial discretization.
• Implemented a Contextual Depth Correction Operator for continuous depth estimation and artifact suppression.
• Proposed the Continuous Geometry Network (CGNet) to jointly optimize spatial and depth discretization.

Main Results:

• CGNet achieved state-of-the-art performance on synthetic and real-world LF datasets.
• Demonstrated significant improvements in accuracy and robustness compared to existing methods.
• Effectively addressed issues of structural ambiguities and depth inaccuracies.

Conclusions:

• UCGR provides a unified approach to handle spatial and depth discretization in LF depth estimation.
• CGNet offers a robust and accurate solution for LF depth estimation challenges.
• The proposed method advances the capabilities of 3D reconstruction, refocusing, and virtual reality applications.