1. What is FastGS?
FastGS is a new, simple, and general 3DGS acceleration framework that can train a scene within 100 seconds while achieving the comparable rendering quality, benefiting from its multi-view consistent densification and pruning strategies. Moreover, it can be easily applied to various challenging tasks such as dynamic scene reconstruction, surface reconstruction, sparse-view reconstruction, large-scale reconstruction and SLAM, achieving 2–7× training acceleration.
2. Training Process Visualization
We demonstrate how FastGS accelerates the training process of different backbones on various scenes.
2.1 Outperform SOTA
FastGS surpasses state-of-the-art Gaussian-based rendering methods in both reconstruction quality and training speed.
Fullscreen the videos to view better.
2.2 Enhancing Baselines
Integrating FastGS into various baseline architectures significantly accelerates their training process.
3. Method Overview
The pipeline of FastGS. (a) We redesign the ADC of the vanilla 3DGS based on multi-view consistency. To accurately assess the importance of each Gaussian, we sample training views and generate the corresponding per-pixel L1 loss maps. For each sampled view, a multi-view score is computed for each Gaussian by counting the number of high-error pixels within its 2D footprint, which is subsequently used to guide Gaussian densification and pruning. (b) Taming-3DGS primarily computes the importance score based on Gaussian-associated properties across sampled views. (c) Speedy-Splat computes the Gaussian score by accumulating Gaussian-associated Hessian approximations across all training views. We visualize the densification results from 0.5K to 15K iterations without any pruning on the far left of the figure, and the pruning results after removing Gaussians at the 6K-th iteration on the far right.
4. FastGS Improving SOTA
