Abstract
A method is disclosed for generating scalable, high-fidelity scene replicas by optimizing 3D Gaussian Splats directly within the equirectangular domain. Conventional pipelines often convert spherical scanner data into perspective views or surface meshes, a process that frequently introduces stitching artifacts, geometric seams, and floaters. To address these limitations, the described approach utilizes a differentiable panorama splat rasterizer. This component processes registered laser scan data and 3D point clouds natively in a spherical format, without projecting imagery onto flat perspective planes. By refining spherical harmonic coefficients and opacity through backpropagation against the original panoramic inputs, the method effectively mitigates conversion-induced distortions. Consequently, robust ground truth depth maps and realistic imagery can be synthesized for diverse device configurations, obviating physical data recapture when adapting to new hardware specifications.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Zhang, Qian; Zhou, Yunwen; Hong, Erin; Turner, Eric Lee; Ji, Dinghuang; and Sturm, Jürgen, "Creating Realistic Virtual Spaces Using Laser Scans and Panoramic Images", Technical Disclosure Commons, (January 29, 2026)
https://www.tdcommons.org/dpubs_series/9248