Organizations: School of Mathematics, Hefei University of Technology, Hefei 230009, China · School of Engineering and Computer Science, Victoria University of Wellington, Wellington 6140, New Zealand · Joint International Research Laboratory of Spatio-temporal Information and Intelligent Location Services, Guilin 541004, China · School of Computer Science and Engineering, University of New South Wales, Sydney 2052, Australia
Sparse-view 3D Gaussian Splatting (3DGS) is prone to overfitting because only a limited number of images are available to constrain scene geometry. Dropout is one regularization strategy for alleviating this problem. We introduce the concept of a lineage family to describe a set of Gaussian primitives derived from the same founder through cloning or splitting operations. The founder Gaussians refer to all Gaussian primitives generated during initialization. Descendants within the same family often jointly contribute to the refinement of adjacent local regions. However, dropping individual Gaussians can be less effective due to the compensation effect from neighboring Gaussians for the dropped ones. To address this issue, we propose Stage-wise Lineage-aware Dropout(SLD), which applies lineage-family dropout during the structural regularization stage and switches to individual dropout during the later refinement stage. We further propose Lineage-Balanced Densification(LBD), which operates during densification by constructing an expanded candidate pool and reallocating the densification budget according to the mean gradient scores and existing descendant sizes of different families. Experiments on the LLFF, MipNeRF360, and Blender datasets show that the proposed framework achieves competitive performance in sparse-view reconstruction
Figures & tables
Figure 1: Overview of the proposed lineage-aware Gaussian optimization method. The input views are first processed through SFM initialization, and each initial Gaussian is assigned a founder identity to form lineage families. The SLD component deactivates descendants from selected families. A currently existing generation-0 Gaussian is excluded from this dropout operation. During densification, LBD uses a lower gradient threshold to build an expanded candidate pool and reallocates the densification budget among different lineage families. SLD controls which Gaussians participate in optimization, while LBD selects densification candidates. Both contribute to the final Gaussian field.
Figure 3: Spatial distribution of lineage families across multiple scenes: The six largest families in each scene are colored, and the remaining Gaussians are shown in gray.
Figure 4: Study of lineage-aware dropout: (a) Probe for all kinds of dropouts. Based on the 10,000 iterations checkpoint of DropGaussian on the LLFF dataset, four removal tests for each LLFF scene are performed: all descendants in the largest lineage family (with count d0), d0 random descendants from top-5 largest families, d0 random Gaussians, and one random descendant from the largest family. The height of bars indicates the PSNR reduction of rendered images before and after one-step of Gaussian removal. (b) Explore the relationship between family dropout and individual Gaussian dropout at the same dropout rate. The height of orange bar indicates the corresponding family dropout rate given the same individual Gaussian dropout rate.
Figure 5: Qualitative comparison on the LLFF dataset. From left to right, the columns show the ground truth, 3DGS, DropoutGS, DropGaussian, and the proposed method. The boxed regions focus on fine structures, object boundaries, and local textures.
Category
Method
3-view
6-view
9-view
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
NeRF-based
Mip-NeRF [ 1 ]
16.11
0.401
0.460
22.91
0.756
0.213
24.88
0.826
0.170
DietNeRF [ 18 ]
14.94
0.370
0.496
21.75
0.717
0.248
24.28
0.801
0.183
RegNeRF [ 30 ]
19.08
0.587
0.336
23.10
0.760
0.206
24.86
0.820
0.161
FreeNeRF [ 40 ]
19.63
0.612
0.308
23.73
0.779
0.195
25.13
0.827
0.160
SparseNeRF [ 35 ]
19.86
0.624
0.328
–
–
–
–
–
–
Table 1: Quantitative comparison on the LLFF dataset. The best results are highlighted in bold, the second-best results are underlined, and the third-best results are marked with dotted underlines.
Figure 6: Qualitative comparison on the MipNeRF360 dataset. From left to right, the columns show the ground truth, 3DGS, DropoutGS, DropGaussian, and the proposed method. The boxed regions highlight thin structures, cluttered background and patterned textures.
Figure 7: Qualitative comparison on the Blender dataset. From left to right, the columns show the ground truth, 3DGS, DropoutGS, DropGaussian, and the proposed method. The boxed regions focus on object contours and appearance details.
Method
12-view
24-view
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
3DGS [ 20 ]
18.52
0.523
0.415
22.80
0.708
0.276
FSGS [ 44 ]
18.80
0.531
0.418
23.70
0.745
0.230
CoR-GS [ 42 ]
19.52
0.558
0.418
23.39
0.727
0.271
DropGaussian [ 31 ]
19.66
0.569
0.374
23.71
0.752
0.227
DropAnSH-GS [ 10 ]
19.93
0.576
0.362
–
–
–
Table 2: Quantitative comparison on the MipNeRF360 dataset. The best results are highlighted in bold, the second-best results are underlined, and the third-best results are marked with dotted underlines.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
3DGS [ 20 ]
21.56
0.847
0.130
DNGaussian [ 24 ]
24.31
0.886
0.088
FSGS [ 44 ]
24.64
0.895
0.095
CoR-GS [ 42 ]
24.32
0.892
0.084
DropGaussian [ 31 ]
25.19
0.890
0.090
DropAnSH-GS [ 10 ]
25.50
0.891
0.088
Table 3: Quantitative comparison on the Blender dataset. The best results are highlighted in bold, the second-best results are underlined, and the third-best results are marked with dotted underlines.
Settings
PSNR ↑
SSIM ↑
LPIPS ↓
SLD
LBD
×
×
19.17
0.646
0.268
×
✓
19.84
0.681
0.213
✓
×
20.50
0.715
0.197
✓
✓
20.70
0.717
0.194
Table 4: Ablation of SLD and LBD under 3-view inputs on the LLFF dataset. The best results are highlighted in bold.
Dropout granularity in stage 2
Founder policy in stage 1 & 2
PSNR ↑
SSIM ↑
LPIPS ↓
Individual
droppable
20.38
0.703
0.205
Individual
preserved
20.22
0.708
0.205
Family
droppable
20.30
0.704
0.204
Family
preserved
20.50
0.715
0.197
Table 5: Ablation of dropout granularity and founder protection policy under 3-view inputs on the LLFF dataset. The best results are highlighted in bold.
ts
PSNR ↑
SSIM ↑
LPIPS ↓
5000
20.4284±0.1145
0.7062±0.0029
0.2029±0.0018
7000
20.4329±0.0519
0.7073±0.0011
0.2018±0.0006
9000
20.4509±0.0123
0.7070±0.0004
0.2032±0.0005
Table 6: Ablation on ts between Stage 2 and Stage 3 under 3-view inputs on the LLFF dataset. The best results are highlighted in bold.
Methods
PSNR ↑
SSIM ↑
LPIPS ↓
FSGS [ 44 ]
20.43
0.682
0.248
FSGS+Ours
20.85
0.719
0.195
CoR-GS [ 42 ]
20.36
0.710
0.202
CoR-GS+Ours
20.72
0.724
0.194
DNGaussian [ 24 ]
19.12
0.591
0.294
DNGaussian+Ours
19.31
0.613
0.287
Table 7: Cross-framework validation under 3-view inputs on the LLFF dataset. The proposed SLD and LBD modules are added to four sparse-view Gaussian Splatting frameworks. The best results are highlighted in bold.
Figure 8: Evolution of lineage-family size concentration during training on LLFF. (a) Family-size Gini coefficient. (b) Percentage of Gaussians belonging to the largest family. SLD and SLD+LBD are compared throughout training.
Figure 9: Final lineage-family size distribution on LLFF. The complementary cumulative distribution is pooled across the eight scenes, and both axes use logarithmic scales. Adding LBD produces a lighter extreme upper tail.
Dropout-based sparse-view 3D Gaussian Splatting (3DGS) methods alleviate overfitting by randomly suppressing Gaussian primitives during training. Existing methods mainly focus on designing increasingly sophisticated dropout strategies, while they overlook the resulting inconsistencies among different dropped Gaussian subsets. This oversight often leads to unstable reconstruction and suboptimal Gaussian representation learning.In this paper, we revisit dropout-based sparse-view 3DGS from a consistency regularization perspective and propose PairDropGS, a Paired Dropout-induced Consistency Regularization framework for sparse-view Gaussian splatting. Specifically, PairDropGS first constructs a pair of the dropped Gaussian subsets from a shared Gaussian field and designs a low-frequency consistency regularization to constrain their low-frequency rendered structures. This design encourages the shared Gaussian field to preserve stable scene layout and coarse geometry under different random dropouts, while avoiding excessive constraints on ambiguous high-frequency details. Moreover, we introduce a progressive consistency scheduling strategy to gradually strengthen the consistency regularization during training for stability and robustness of reconstruction. Extensive experiments on widely-used sparse-view benchmarks demonstrate that PairDropGS achieves superior training stability, significantly outperforms existing dropout-based 3DGS methods in reconstruction quality, while exhibiting the simplicity and plug-and-play nature for improving dropout-based optimization.
Hantang Li, Qiang Zhu, Xiandong Meng +3
School of Computer Science, Harbin Institute of Technology, Shenzhen, China · Pengcheng Laboratory, Shenzhen, China · Smart Coding Institute, Pengcheng Laboratory, Shenzhen, China +1
Recent 3D Gaussian Splatting (3DGS) Dropout methods address overfitting under sparse-view conditions by randomly nullifying Gaussian opacities. However, we identify a neighbor compensation effect in these approaches: dropped Gaussians are often compensated by their neighbors, weakening the intended regularization. Moreover, these methods overlook the contribution of high-degree spherical harmonic coefficients (SH) to overfitting. To address these issues, we propose DropAnSH-GS, a novel anchor-based Dropout strategy. Rather than dropping Gaussians independently, our method randomly selects certain Gaussians as anchors and simultaneously removes their spatial neighbors. This effectively disrupts local redundancies near anchors and encourages the model to learn more robust, globally informed representations. Furthermore, we extend the Dropout to color attributes by randomly dropping higher-degree SH to concentrate appearance information in lower-degree SH. This strategy further mitigates overfitting and enables flexible post-training model compression via SH truncation. Experimental results demonstrate that DropAnSH-GS substantially outperforms existing Dropout methods with negligible computational overhead, and can be readily integrated into various 3DGS variants to enhance their performances. Project Website: https://sk-fun.fun/DropAnSH-GS
Shuangkang Fang, I-Chao Shen, Xuanyang Zhang +5
Beihang University · The University of Tokyo · StepFun
Novel view synthesis from sparse inputs remains challenging for 3D Gaussian Splatting (3DGS) due to ambiguous geometry, cross-view inconsistency, and missing details in under-constrained regions, resulting in degraded reconstruction and unstable rendering. To tackle these issues, we propose D3GS, a Depth-DINO-Diffusion guided sparse-view Gaussian reconstruction framework that jointly enhances geometry and appearance. D3GS first recovers a high-resolution, metric depth map via diffusion-based completion and DPT (Dense Prediction Transformer) refinement, providing robust Gaussian initialization and geometric constraints. Then, a DINO-guided view-consistent learning is introduced to augment Gaussian attributes with structural features, improving multi-view consistency. Finally, a diffusion-based Gaussian refinement module injects generative priors into an iterative optimization strategy, enhancing high-frequency geometric and appearance details within the Gaussian representation. Experiments on DTU, LLFF, and Mip-NeRF 360 show that D3GS achieves consistent and substantial improvements over strong baselines, with ablation studies validating the effectiveness and complementary roles of each component.
Yunqi Gao, Zhanfeng Liao, Hanzhang Tu +7
Central China Normal University · Tsinghua University · Fuzhou University +4