Structure-from-motion in micro-image domain for uncalibrated plenoptic 2.0 cameras.

Area of Science:

• Computer Vision
• Photogrammetry
• Computational Imaging

Background:

• Plenoptic cameras capture depth information using a micro-lens array, offering advantages over traditional monocular cameras.
• Classical structure-from-motion (SfM) methods face challenges with low angular disparity and scale ambiguity, especially with complex camera setups.

Purpose of the Study:

• To introduce a novel structure-from-motion (SfM) method tailored for raw micro-images from plenoptic 2.0 cameras.
• To address limitations of traditional SfM by leveraging inherent disparity information in plenoptic imaging.

Main Methods:

• Developed a SfM pipeline that identifies pinhole camera constraints and utilizes plenoptic camera disparity.
• Enabled robust initialization and reconstruction without calibration patterns or subaperture view extraction.
• Reconstructed scenes from multiple uncalibrated plenoptic cameras.

Main Results:

• Achieved 10% error accuracy in relative pose estimation on natural and synthetic datasets, comparable to calibration-based methods.
• Demonstrated robustness to coarse initialization and ability to reconstruct scenes with parallel cameras.
• Outperformed reconstruction methods based on pinhole camera conversion.

Conclusions:

• The proposed SfM method effectively reconstructs 3D scenes from uncalibrated plenoptic cameras.
• It overcomes key limitations of classical SfM, including scale and angular disparity ambiguity.
• Offers a more accurate and versatile solution for 3D scene reconstruction using plenoptic imaging.