3D Gaussian Splatting (3DGS) has emerged as a powerful technique for novel view synthesis, demonstrating remarkable capability in high-fidelity scene reconstruction through its Gaussian primitive representations. However, the computational overhead induced by the massive number of primitives poses a significant bottleneck to training efficiency.
To overcome this challenge, we propose Group Training, a simple yet effective strategy that organizes Gaussian primitives into manageable groups, optimizing training efficiency and improving rendering quality. This approach shows universal compatibility with existing 3DGS frameworks, including vanilla 3DGS and Mip-Splatting, consistently achieving accelerated training while maintaining superior synthesis quality. Extensive experiments reveal that our straightforward Group Training strategy achieves up to 30% faster convergence and improved rendering quality across diverse scenarios.
Optimizing all Gaussian primitives concurrently during training is not necessary. Group Training involves periodically dividing all Gaussian primitives. Specifically, at regular iteration intervals, Gaussians from all groups are merged before rendering the training view. Subsequently, all Gaussian primitives are categorized into the Under-training Group and the Caching Group according to a specified sampling strategy. Before the next grouping, the Under-training Group is utilized for gaussian densification (Iteration 0~15K) or optimization (Iteration 15~30K), while the Caching Group remains inactive and does not participate in any calculations.
Sampling strategy is crucial for accelerating training in 3D Gaussian Splatting. Through both formula derivation and experimental validation, we demonstrate the acceleration efficacy of the Opacity-based Prioritized Sampling Strategy (OPS). Specifically, OPS enables more Effective Densification and more Efficient Rendering. Adjusting the slider allows you to view the opacity distribution of the Gaussian primitives selected for densification and all Gaussian primitives. The comparison illustrates the critical role of high-opacity Gaussian primitives in densification.
Gaussian primitives with higher opacity serve as the primary contributors to densification of 3DGS. \[ \mathbb{E} \left[\frac{\partial \hat{C}}{\partial \alpha_m} \right] = \frac{(c_0-c_\mathrm{bg}) T_{\mathrm{saturation}} }{1-\mathbb{E}[o_m] \cdot \mathbb{E}[G_m^\mathrm{2D}]} \]
Gaussian primitives with higher opacity enable faster rendering through faster achieving \(\alpha\) saturation. \[ \begin{aligned} \mathbb{E}[T_{N}] &= (1-\mathbb{E}\left[{\alpha_i}\right])^{N} \\ &= (1-\mathbb{E}[o_i] \cdot \mathbb{E}[G_i^\mathrm{2D}])^N \end{aligned} \]
Group Training achieves faster reconstruction speeds by controlling the training of fewer Gaussian primitives, without compromising reconstruction quality. The blue pixels represent the projected positions of Gaussians onto the image plane. Compared to the Baseline, Group Training exhibits a significantly lower density of Gaussian primitives. Consequently, the number of hitten Gaussians per optimization iteration is substantially reduced compared to the Baseline, leading to faster rendering speeds and higher FPS. For specific metric comparisons, refer to the Hitten Gaussians - FPS scatter plot. Adjusting the slider at the bottom to view comparisons at different training iterations.
@article{wang2024faster,
title={Faster and better 3d splatting via group training},
author={Wang, Chengbo and Ma, Guozheng and Xue, Yifei and Lao, Yizhen},
journal={arXiv preprint arXiv:2412.07608},year={2024}
}