| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 14.39 MB | Adobe PDF |
Abstract(s)
This study investigates the integration of multi-lens setups,
specifically wide and ultrawide lenses, into Gaussian Splatting
frameworks to enhance 3D scene reconstruction. By leveraging
the fixed extrinsic relationship between the lenses, the proposed
approach combines multiple RGB perspectives to improve depth
accuracy, geometric precision, and texture fidelity. Unlike
traditional single-lens approaches, this method utilizes the
expanded field of view provided by the ultrawide lens as
auxiliary data to complement the wide lens. No particular
benefit was presented from using the specific approach this
study (with limitations accounted for in the full text), yet during
testing, reconstructions that used solely the ultrawide lens scored
very competitive metrics, specially given then their sample was
lower both image count and image resolution, important
variables for such reconstructions. This work highlights the
potential of implementing a native fisheye camera model in the
state-of-the-art Gaussian Splatting pipeline.
Description
Keywords
Computer vision Gaussian Splatting Radiance Fields NeRFs Volumetric reconstruction
