Dropping Anchor and Spherical Harmonics for Sparse-view Gaussian Splatting

Shuangkang Fang; I-Chao Shen; Xuanyang Zhang; Zesheng Wang; Yufeng Wang; Wenrui Ding; Gang Yu; Takeo Igarashi

cs.CVFeb 24, 2026

Dropping Anchor and Spherical Harmonics for Sparse-view Gaussian Splatting

Authors: Shuangkang Fang, I-Chao Shen, Xuanyang Zhang, Zesheng Wang, Yufeng Wang, Wenrui Ding, Gang Yu, Takeo Igarashi

Organizations: Beihang University · The University of Tokyo · StepFun

Abstract

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

Figures & tables

Figure 1: Difference between DropGaussian [ 46 ] and our DropAnSH-GS . (a) DropGaussian randomly sets the opacity of individual Gaussians to zero. (b) DropAnSH-GS selects random Gaussians as anchors, discards their neighbors, and applies Dropout to high-degree SH. Anchor-based Dropout eliminates contiguous 3D regions, which effectively inhibits adjacent Gaussians from compensating for the dropped ones, forcing remaining Gaussians to learn more robust scene representations, thereby strengthening the Dropout regularization effect. Dropping SH further reduces overfitting and yields a more compact model.

Figure 2: Compensation effect in neighbor Gaussians . (a) We measure the spatial autocorrelation of opacity and color between Gaussians at varying distances using Moran’s I metric [ 40 , 37 ] . Closer Gaussians exhibit higher similarity in opacity and color. This spatial redundancy implies that dropping only individual Gaussians can be easily compensated for by their surrounding similar neighbors, thereby limiting the effectiveness of dropout. (b) We report the Mean Absolute Error between rendered images before and after applying two different Dropout strategies, S1 and S2. When dropping a similar number of Gaussians, the strategy that drops Gaussians and their neighboring Gaussians (S1) has a larger impact on rendering quality, as it avoids simple compensation by neighboring Gaussians. (c) In regions surrounding the dropped Gaussians, the S1 strategy activates a greater number of remaining Gaussians, resulting in stronger gradient updates.

Figure 3: Overfitting caused by high-degree SH . On the LLFF dataset, (a) under full-view conditions, moderately increasing the degree of spherical harmonics can improve the performance of the 3DGS model. However, (b) under sparse-view settings, using high-degree spherical harmonics leads to performance degradation and a significant increase in model size, indicating that spherical harmonics themselves also constitute a source of overfitting.

Figure 4: Overview of DropAnSH-GS . During the training phase, we randomly select a set of anchor Gaussians and drop these anchors and their surrounding neighbours within a local spatial region, thereby reducing the likelihood of neighboring Gaussians compensating for one another and enhancing the regularization effect. Simultaneously, high-degree SH coefficients are randomly dropped to further strengthen regularization and yield a more compact model representation.

Figure 5: Qualitative comparison on the LLFF dataset (3 views).

	Methods	3-view			6-view			9-view
	Methods	PSNR↑	SSIM↑	LPIPS↓	PSNR↑	SSIM↑	LPIPS↓	PSNR↑	SSIM↑	LPIPS↓
NeRF-based	Mip-NeRF [ 4 ]	16.11	0.401	0.460	22.91	0.756	0.213	24.88	0.826	0.170
	DietNeRF [ 22 ]	14.94	0.370	0.496	21.75	0.717	0.248	24.28	0.801	0.183
	RegNeRF [ 42 ]	19.08	0.587	0.336	23.10	0.760	0.206	24.86	0.820	0.161
	FreeNeRF [ 60 ]	19.63	0.612	0.308	23.73	0.779	0.195	25.13	0.827	0.160
3DGS-based	3DGS [ 25 ]	19.17	0.646	0.268	23.74	0.807	0.162	25.44	0.860	0.096

Table 1: Quantitative comparison on the LLFF dataset (3, 6 and 9 views).

Figure 6: Qualitative comparisons on the Blender (8 views) and MipNeRF-360 (12 views) datasets.

	MipNeRF-360		Blender
Methods	PSNR↑	Size (Mb)↓	PSNR↑	Size (Mb)↓
3DGS [ 25 ]	18.58	143.4	22.13	6.5
DNGaussian [ 28 ]	18.84	80.7	24.31	5.5
FSGS [ 66 ]	18.80	172.1	24.64	8.4
CoR-GS [ 63 ]	19.42	117.4	24.31	6.3
DropoutGS [ 58 ]	19.17	68.4	24.79	5.2

Table 2: Quantitative comparison on the MipNeRF-360 (12 views) and Blender (8 views) datasets . SH n denotes retaining only the first n degrees of SH coefficients during inference.

Figure 7: Compatibility to different 3DGS variants on the LLFF dataset (3 views) . The proposed DropAnSH-GS can be easily integrated into other 3DGS-based methods and improve their performance.

Methods	PSNR↑	SSIM↑	LPIPS↓
FSGS [ 66 ]	20.43	0.682	0.248
FSGS+Ours	20.72	0.713	0.205
CoR-GS [ 63 ]	20.36	0.710	0.202
CoR-GS+Ours	20.74	0.718	0.203
DNGaussian [ 28 ]	19.12	0.591	0.294
DNGaussian+Ours	19.71	0.639	0.268

Table 3: Compatibility to 3DGS variants on the LLFF dataset (3 views) . The proposed DropAnSH-GS can be easily integrated into various 3DGS-based methods and improve their performance.

Method	PSNR↑	SSIM↑	LPIPS↓	Total Time↓
Method	LLFF dataset (3 views)
3DGS	19.17	0.646	0.268	741.6s
Ours	20.68	0.724	0.194	760.2s
	Blender dataset (8 views)
3DGS	22.13	0.855	0.132	863.3s
Ours	25.50	0.891	0.088	887.7s

Table 4: Comparison of training time on different datasets. We achieve a substantial performance improvement with a negligible increase in computational time.

$p_{a}$	PSNR↑	SSIM↑	LPIPS↓	$k$	PSNR↑	SSIM↑	LPIPS↓	$p_{sh}$	PSNR↑	SSIM↑	LPIPS↓
0.005	20.25	0.705	0.211	1	20.12	0.701	0.216	0.1	20.56	0.721	0.198
0.01	20.50	0.718	0.194	5	20.39	0.719	0.200	0.2	20.68	0.724	0.194
0.02	20.68	0.724	0.194	10	20.68	0.724	0.194	0.3	20.65	0.725	0.197
0.03	20.32	0.710	0.202	15	20.43	0.721	0.198	0.4	20.44	0.713	0.202
0.04	19.97	0.665	0.257	20	20.07	0.706	0.224	0.5	20.49	0.711	0.208

Table 5: Hyperparameter sensitivity of the anchor selection probability

p_{a}

, number of neighbors

k

, and SH Dropout probability

p_{sh}

. The evaluation is conducted on the LLFF dataset (3 views).

Settings		PSNR↑	SSIM↑	LPIPS↓
Drop Anchor	Drop SH	PSNR↑	SSIM↑	LPIPS↓
$\times$	$\times$	19.17	0.646	0.268
$\checkmark$	$\times$	20.47	0.713	0.200
$\times$	$\checkmark$	19.59	0.641	0.247
$\checkmark$	$\checkmark$	20.68	0.724	0.194

Table 6: Ablation results of the designed components on the LLFF datasets (3 views).

	PSNR↑	SSIM↑	LPIPS↓
Drop SH Randomly	25.12	0.885	0.097
Drop SH by Degree (Ours)	25.50	0.891	0.088

Table 7: Ablation results of different SH Dropout strategies on the Blender dataset (8 views).

	PSNR↑	SSIM↑	LPIPS↓
w/o - SH 0	24.57	0.879	0.112
w/ - SH 0	25.04	0.883	0.093
w/o - SH 1	24.96	0.885	0.096
w/ - SH 1	25.34	0.890	0.090
w/o - SH 2	25.21	0.887	0.091
w/ - SH 2	25.47	0.891	0.089

Table 8: Effect of with and without Drop SH strategy on the Blender dataset (8 views). The Drop SH strategy encourages the effective parameters to concentrate in lower-degree SH coefficients, thereby mitigating the performance degradation caused by direct SH truncation.

Figure 8: Visualization of the entire reconstruction scene. Our method produces results that are more complete and natural.

Methods	PSNR↑	SSIM↑	LPIPS↓	Size (MB)↓
3DGS	18.58	0.526	0.419	143.4
DNGaussian	18.84	0.543	0.468	80.7
FSGS	18.80	0.531	0.418	172.1
CoR-GS	19.42	0.556	0.418	117.4
DropoutGS	19.17	0.558	0.386	68.4
DropGaussian	19.66	0.569	0.374	120.7

Table 9: Quantitative comparison on the MipNeRF-360 dataset (12 views) . SH n denotes retaining only the first n degrees of SH coefficients during inference.

Methods	PSNR↑	SSIM↑	LPIPS↓	Size (Mb)↓
3DGS	22.13	0.855	0.132	6.5
DNGaussian	24.31	0.886	0.088	5.5
FSGS	24.64	0.895	0.095	8.4
CoR-GS	24.31	0.886	0.091	6.3
DropoutGS	24.79	0.877	0.110	5.2
DropGaussian	25.17	0.882	0.100	6.0

Table 10: Quantitative comparison on the Blender dataset (8 views) . SH n denotes retaining only the first n degrees of SH coefficients during inference.

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 1,013 papers

PairDropGS: Paired Dropout-Induced Consistency Regularization for Sparse-View Gaussian Splatting
May 12, 2026Hantang Li, Qiang Zhu, Xiandong Meng +33D Gaussian Splatting Sparse-View 3D Reconstruction
School of Computer Science, Harbin Institute of Technology, Shenzhen, China · Pengcheng Laboratory, Shenzhen, China · Smart Coding Institute, Pengcheng Laboratory, Shenzhen, China +1
Anchor3DGS: Feed-Forward 3D Gaussian Splatting with Compact Anchor-Based Representation
Oct 15, 2026Sheng Ye, Jiangke Lin, Fudong Wang +33D Gaussian Splatting Multi-View 3D Reconstruction
Tsinghua University · Ant Group, Inclusion AI
Compressing 3D Gaussian Splatting via Cross-Representation Priors
Sep 19, 2026Yezheng Zhang, Huanxiong Liang, Chuqin Zhou +23DGS Compression Gaussian Splatting Compression
Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University, Shanghai 200240, China

May 12, 2026J/K move · Enter open · S save

PairDropGS: Paired Dropout-Induced Consistency Regularization for Sparse-View Gaussian Splatting

Hantang Li, Qiang Zhu, Xiandong Meng, Xingtao Wang+2

School of Computer Science, Harbin Institute of Technology, Shenzhen, China · Pengcheng Laboratory, Shenzhen, China · Smart Coding Institute, Pengcheng Laboratory, Shenzhen, China+1

3D Gaussian Splatting Sparse-View 3D Reconstruction Consistency Regularization

Open paper AskPDF arXiv

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.

Abstract

3D Gaussian Splatting

Explore similar work

PairDropGS: Paired Dropout-Induced Consistency Regularization for Sparse-View Gaussian Splatting

Anchor3DGS: Feed-Forward 3D Gaussian Splatting with Compact Anchor-Based Representation

Compressing 3D Gaussian Splatting via Cross-Representation Priors