Iterative Occlusion-Aware Light Field Depth Estimation Using 4-D Geometrical Cues.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Computational Imaging

Background:

• Light field cameras capture 4D light field geometry, embedding 3D scene information.
• Existing depth estimation methods often rely on gradient information, optimization, or learning-based approaches.

Purpose of the Study:

• To develop a novel, non-learning-based method for light field depth estimation.
• To explicitly utilize 4D geometric cues for improved depth accuracy.

Main Methods:

• A novel 4D geometric model of the light field is proposed.
• The model explicitly considers surface normal accuracy and occlusion regions.
• Depth/disparity is estimated by analyzing intersections of 2D planes in 4D space.

Main Results:

• The proposed method achieves a 26.3% lower Median Angle Error on planar surfaces compared to state-of-the-art.
• The method demonstrates competitive performance in terms of Mean Squared Error (MSE) and Bad Pixel (Badpix) metrics.
• Outperforms both learning-based and non-learning-based state-of-the-art methods in surface normal accuracy.

Conclusions:

• The novel 4D geometric model offers a more explainable and accurate approach to light field depth estimation.
• Explicitly leveraging 4D geometric cues enhances surface normal accuracy.
• The method provides a competitive alternative to current state-of-the-art depth estimation techniques.