Robust Light Field Depth Estimation Using Occlusion-Noise Aware Data Costs.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Depth estimation is crucial for light field applications.
• Existing methods struggle with noisy and occluded scenes.

Purpose of the Study:

• To develop a light field depth estimation method resilient to noise and occlusion.
• To improve depth map accuracy in challenging visual environments.

Main Methods:

• Proposed two novel data costs: constrained angular entropy cost (CAE) and constrained adaptive defocus cost (CAD).
• CAE mitigates occluder and noise effects in angular patches.
• CAD provides low cost in occluded regions and noise robustness.
• Integrated CAE and CAD for enhanced occlusion and noise invariance.
• Applied cost volume filtering and graph cut optimization for depth map refinement.

Main Results:

• The integrated data costs significantly enhance robustness against occlusion and noise.
• Experimental results demonstrate high-quality depth map generation across diverse scenes.
• The proposed method outperforms existing state-of-the-art techniques in qualitative and quantitative evaluations.

Conclusions:

• The novel method offers superior occlusion and noise invariant capability for light field depth estimation.
• It achieves state-of-the-art performance, producing accurate depth maps from complex scenes.