A Lightweight Depth Estimation Network for Wide-Baseline Light Fields.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Traditional and Convolutional Neural Network (ConvNet)-based methods for light field depth estimation are primarily limited to narrow-baseline scenarios.
• Wide-baseline light fields present a challenging yet promising area for depth prediction due to their potential for richer scene information.

Purpose of the Study:

• To explore the feasibility and capability of ConvNets for depth estimation in wide-baseline light fields.
• To develop a lightweight, end-to-end trained convolutional network for accurate depth inference from wide-baseline light fields.

Main Methods:

• A large-scale, diverse synthetic wide-baseline dataset with labeled data was created to address the scarcity of training samples.
• A novel lightweight convolutional network, LLF-Net, was designed, incorporating a cost volume for variable angular inputs and an attention module for occlusion handling.

Main Results:

• LLF-Net demonstrated superior performance on both synthetic and real-world wide-baseline light field datasets compared to state-of-the-art methods.
• The network also showed performance improvements when evaluated on existing narrow-baseline datasets, highlighting its versatility.

Conclusions:

• LLF-Net effectively addresses the challenges of depth estimation in wide-baseline light fields.
• The proposed network offers a practical solution for real-world applications and enhances existing depth estimation techniques.