We're beginning to see the fruits of nerfstudio's efforts with the Splatfacto model, evolving into a new methodology with the introduction of OpenSplat. OpenSplat is a free and open-source implementation of 3D Gaussian Splatting, written in C++. It focuses on portability, efficiency, and speed.
OpenSplat takes camera poses and sparse points and creates a .ply file that can be used for viewing, editing and rendering in an ever growing list of platforms that support Gaussian Splatting.
Splatfacto represents nerfstudio's entirely in-house implementation of Gaussian Splatting and is licensed under Apache 2.0. This license facilitates further developments like OpenSplat. Moreover, OpenSplat adopts the AGPLv3 license, enabling commercial use.
Piero Toffanin, the developer behind OpenSplat, has set ambitious goals aimed at broadening its accessibility and utility. Among these objectives is the adaptation for AMD and Apple devices, allowing users to train models on their own equipment without the necessity of CUDA. This would greatly expand the potential user base of radiance fields.
Other details include an effort to boost performance. Because OpenSplat is built with C++, there's no interpreter lock and it will be easier to optimize memory usage and speed. That said, it includes initiatives to improve speed and reduce memory usage, ensuring that the tool is both faster and more resource-efficient. Additionally, they are looking to implement distributed computation capabilities. This will allow multiple machines to collaborate on computations, significantly accelerating processing times and enabling more complex projects. The roadmap also includes plans for compressed scene outputs, aiming to make the storage and sharing of projects more manageable without compromising on detail or quality.
At present, running it requires CUDA and an NVIDIA GPU. The full list of dependencies includes CUDA, OpenCV, and Libtorch. OpenSplat currently supports Ubuntu 20.04 and Windows, making it a versatile tool for developers.
This is an extremely exciting addition to the radiance field community and I am excited to watch it continue to grow. For more information and updates, visit Piero's GitHub page.
New to Gaussian Splatting? Start here
