Depth Estimation from Light Field Geometry Using Convolutional Neural Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Computational Imaging

Background:

• Light field imaging offers advanced depth estimation beyond traditional stereo or monocular methods.
• Existing methods struggle to balance computational efficiency with depth estimation accuracy.
• Deep learning algorithms can leverage abundant light field data for improved depth estimation.

Purpose of the Study:

• To develop a novel deep learning model for efficient and accurate depth estimation from light field images.
• To explore the use of Epipolar Plane Images (EPIs) as input for enhancing depth estimation.
• To achieve a balance between computational time and accuracy in light field depth estimation.

Main Methods:

• Designed ESTNet, a convolutional neural network (CNN) with a three-input stream architecture (horizontal EPI, vertical EPI, central view image).
• Utilized EPI synthetic images derived from light field data to improve feature extraction.
• Employed an encoding-decoding structure with skip-connections to fuse local and semantic features.
• Trained and tested the model on synthetic and real-world light field datasets.

Main Results:

• ESTNet demonstrated a reasonable architecture through ablation experiments.
• The model successfully balanced depth estimation accuracy and computational time on both synthetic and real datasets.
• EPI-based inputs significantly enhanced feature extraction and depth estimation performance.

Conclusions:

• ESTNet provides an effective solution for fast and accurate depth estimation in light field imaging.
• The proposed method successfully integrates geometric cues from EPIs with CNN capabilities.
• This approach offers a promising direction for real-time applications requiring precise depth information.