Cube-Splat Introduces Open-Source Panoramic Gaussian Splatting SLAM

A paper on Cube-Splat has been published on arXiv — a method for simultaneous localization and mapping (SLAM) based on 3D Gaussian Splatting for panoramic 360-degree images. The authors split each panoramic frame into a cubemap of four virtual perspective views sharing a common optical center, which allows consistently updating a single pose state from observations across all faces. To evaluate the method, they present SynPano, a synthetic dataset with photorealistic scenes and parameterized trajectories. The details are laid out in the paper on arXiv.

What it means

Cube-Splat advances the direction that Gaussian Splatting (3DGS) belongs to in our reference. The base 3DGS method from Inria and MPII is an open source 3D generator that trains in a PyTorch environment with CUDA extensions and requires 24 GB of VRAM to reach the quality claimed in the original paper. The license of the base code permits non-commercial research purposes only; commercial use requires explicit consent from the licensors. Cube-Splat differs from base 3DGS in its problem statement: it is not offline model generation from a set of shots, but simultaneous camera localization and mapping in real time, and specifically for panoramic cameras. The authors have released the source code and the SynPano dataset publicly, but the paper's description does not state the VRAM requirements, the launch platform, or the license of the Cube-Splat code. For a reader looking for a tool for their task, this means one more open method has appeared in the Gaussian Splatting family, but with a different scope — panoramic SLAM rather than static reconstruction. It is too early to judge practical applicability on local hardware: the authors provide no information about resource consumption or the terms of code distribution.