Fast and Accurate 3D Measurement Based on Light-Field Camera and Deep Learning

Area of Science:

• Computer Vision
• 3D Reconstruction
• Optical Engineering

Background:

• Light-field cameras capture angular and spatial light information, enabling depth and disparity calculation from a single image.
• This capability facilitates the creation of 3D measurement systems and 3D model recovery.
• Existing disparity estimation algorithms struggle with reflective and texture-less regions in light-field images.

Purpose of the Study:

• To develop a novel method for accurate disparity estimation in light-field images.
• To address the challenges posed by reflective and texture-less areas.
• To improve 3D reconstruction accuracy using light-field data.

Main Methods:

• Introduction of VommaNet, an end-to-end deep learning network.
• Utilizing multi-scale feature retrieval specifically for challenging regions.
• Evaluation on both synthetic and real-world light-field datasets.

Main Results:

• VommaNet achieves accurate disparity estimation, particularly in reflective and texture-less areas.
• The network demonstrates comparable or superior performance to state-of-the-art algorithms in various regions.
• Successful application to both synthetic and real-world light-field image data.

Conclusions:

• VommaNet effectively overcomes limitations in existing disparity estimation techniques for light-field cameras.
• The proposed network offers a robust solution for 3D measurement and reconstruction.
• This advancement has implications for applications requiring precise 3D information from single images.