Multi-Viewpoint Panorama Construction With Wide-Baseline Images.

Area of Science:

• Computer Vision
• Image Processing
• Computational Geometry

Background:

• Traditional image stitching methods struggle with wide baselines and non-planar scenes.
• Existing techniques often rely on complex 3D reconstruction, limiting their applicability.
• Achieving both local perspective and global orthogonality in panoramas is challenging.

Purpose of the Study:

• To develop a novel image stitching approach for generating visually plausible panoramic images.
• To enable stitching of images with wide baselines and non-planar scene structures.
• To optimize alignment and regularity in 2D without resorting to 3D reconstruction.

Main Methods:

• A mesh-based framework is designed to optimize alignment and regularity in 2D.
• A global objective function, incorporating alignment and prior constraints, is solved.
• The method constructs panoramic images that are locally perspective and globally nearly orthogonal.

Main Results:

• The proposed approach successfully produces visually plausible panoramic images.
• It demonstrates effectiveness in handling wide baselines and non-planar scenes.
• Improved composition and performance in misaligned areas were observed in experiments.

Conclusions:

• The novel mesh-based framework offers an effective solution for image stitching.
• The method overcomes limitations of previous approaches regarding baseline and scene complexity.
• Experimental results validate the technique's robustness on challenging datasets.