DropoutGS: Dropping Out Gaussians for Better Sparse-view Rendering

TL;DR

This work tackles the under-constrained problem of sparse-view novel view synthesis with 3D Gaussian Splatting (3DGS) by addressing overfitting through Random Dropout Regularization (RDR) and enhancing detail via Edge-guided Splitting Strategy (ESS). RDR regularizes training by randomly dropping Gaussians, effectively creating an ensemble of low-complexity sub-models, while ESS splits large Gaussian primitives in high-edge regions to recover high-frequency details. The combined DropoutGS framework is shown to be plug-in compatible with diverse 3DGS methods and achieves state-of-the-art results on benchmarks such as Blender, LLFF, and DTU, with improved geometry and fewer artifacts. Together, these strategies yield smoother renderings and better depth maps under sparse views, highlighting a practical route to robust sparse-view 3D scene reconstruction.

Abstract

Although 3D Gaussian Splatting (3DGS) has demonstrated promising results in novel view synthesis, its performance degrades dramatically with sparse inputs and generates undesirable artifacts. As the number of training views decreases, the novel view synthesis task degrades to a highly under-determined problem such that existing methods suffer from the notorious overfitting issue. Interestingly, we observe that models with fewer Gaussian primitives exhibit less overfitting under sparse inputs. Inspired by this observation, we propose a Random Dropout Regularization (RDR) to exploit the advantages of low-complexity models to alleviate overfitting. In addition, to remedy the lack of high-frequency details for these models, an Edge-guided Splitting Strategy (ESS) is developed. With these two techniques, our method (termed DropoutGS) provides a simple yet effective plug-in approach to improve the generalization performance of existing 3DGS methods. Extensive experiments show that our DropoutGS produces state-of-the-art performance under sparse views on benchmark datasets including Blender, LLFF, and DTU. The project page is at: https://xuyx55.github.io/DropoutGS/.

Figures (12)

• Figure 1: Results produced by 3DGS with different numbers of Gaussians. The training loss curves and rendered results are visualized for comparison. Compared with the model with 1k Gaussians, the one with 10k Gaussians suffers overfitting and produces inferior results.
• Figure 2: An overview of our framework. RDR is first employed to alleviate the overfitting issue. Then, ESS is adopted to split the large Gaussians to better capture high-frequency details.
• Figure 3: The relationship between the amount of training data and model complexity. We investigate the performance of 3DGS with different primitive settings under varying numbers of sparse views. The results show that models achieve optimal performance when their complexity matches the training data.
• Figure 4: The scale distribution of the Gaussians learned by models with different complexities. The models with 10k and 20k primitives have a large portion of small-scale Gaussians. In contrast, the model with 1k primitives obtains more Gaussians with larger scales.
• Figure 5: Visualization of the gradient maps. The dropouted Gaussian is annotated with a red circle. Thus brighter regions correspond to higher gradients.