UGOD: Uncertainty-Guided Opacity and Dropout for Sparse-View 3D Gaussian Splatting

Zhihao Guo; Peng Wang; Zidong Chen; Xiangyu Kong; Yan Lyu; Guanyu Gao; Chenghao Qian; Ziyang Wang

cs.CVSep 30, 2026

UGOD: Uncertainty-Guided Opacity and Dropout for Sparse-View 3D Gaussian Splatting

Authors: Zhihao Guo, Peng Wang, Zidong Chen, Xiangyu Kong, Yan Lyu, Guanyu Gao, Chenghao Qian, Ziyang Wang, +2 more

Organizations: Manchester Metropolitan University, Manchester, United Kingdom. · University of Surrey, Guildford, United Kingdom. · Imperial College London, London, United Kingdom. · University of Exeter, Exeter, United Kingdom. · Southeast University, Nanjing, China. · Nanjing University of Science and Technology, Nanjing, China. · University of Leeds, Leeds, United Kingdom. · Aston University, Birmingham, United Kingdom.

Abstract

Sparse-view 3D Gaussian Splatting is prone to overfitting because limited observations leave many Gaussian primitives weakly constrained, yet their contributions are still accumulated through alpha blending. Without uncertainty estimation, the renderer cannot distinguish unreliable primitives from well-constrained ones, allowing their erroneous contributions to corrupt novel-view synthesis. We introduce UGOD, an uncertainty-guided framework that estimates a view-dependent uncertainty score for each Gaussian and uses it to regulate its rendering contribution. A lightweight uncertainty head conditioned on Gaussian attributes and viewing direction predicts this score, which then drives a differentiable opacity-modulation mechanism that attenuates high-uncertainty primitives before compositing. During training, a detached soft-dropout branch applies an uncertainty-controlled continuous keep mask to discourage the model from relying on poorly constrained Gaussians and thereby reduce overfitting. Crucially, detaching the uncertainty score prevents gradients from this stochastic regulariser from biasing or collapsing the uncertainty prediction. Experiments on Mip-NeRF~360 and LLFF show that UGOD improves sparse-view novel-view synthesis while producing more compact Gaussian representations than the compared methods. These results demonstrate that Gaussian uncertainty provides an effective rendering-time control for sparse-view reconstruction.

Figures & tables

Figure 1: Average PSNR and final-Gaussian reduction versus 3DGS* across MiP-NeRF 360 (24/8 views) and LLFF (3 views). Higher is better on every axis; UGOD leads all six.

Figure 2: Overview of UGOD. Sparse input views are reconstructed by SfM and initialise 3D Gaussian primitives (left). For each visible Gaussian, the uncertainty head combines the viewing direction, HashGrid position encoding, geometric attributes, and compressed SH appearance to predict a view-dependent uncertainty $u_{i}$ (lower left). This score drives two gradient-decoupled mechanisms (centre): normalised uncertainty differentiably gates opacity before compositing, whereas its detached copy determines the training-only Concrete soft-dropout probability. The resulting Gaussians are jointly optimised using supervised depth, pseudo-view depth, and photometric losses (right). At inference, soft dropout and relative-uncertainty gating are disabled, and deterministic raw-uncertainty opacity modulation is used.

Setting	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	#Gaussians	vs 3DGS*
LLFF (3-view)	3DGS*	14.6543	0.4379	0.3990	254,350	–
	DropGS	14.6443	0.4557	0.3894	210,414	1.2 $\times$
	FSGS (w/ depth)	19.0457	0.6230	0.2514	151,173	1.7 $\times$
	Ours (full)	19.3300	0.6370	0.2449	102,280	2.5 $\times$
Mip360 (8-view)	3DGS*	11.9591	0.2747	0.6095	864,393	–
	DropGS	12.7267	0.3193	0.6026	716,133	1.2 $\times$

Table 1: Scene-average results on LLFF (3 views) and MiP-NeRF 360 (8/24 views). Reduction is relative to 3DGS*; best and second-best values are bold and underlined.

Setting	Variant	$N$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	#Gaussians $\downarrow$
LLFF (3-view)	Ours	7	18.2386	0.5989	0.2701	180,905
	w/o soft dropout	7	17.9529	0.5700	0.2877	183,460
	Ours (full)	7	19.3300	0.6370	0.2449	102,280
Mip360 (8-view)	w/o soft dropout	5	13.9438	0.3672	0.6074	108,008
	w/ soft dropout	5	12.9741	0.3475	0.6267	102,644
	Ours (full)	6	14.1435	0.3682	0.6179	95,662

Table 2: Scene-average component ablation.

N

is the number of scenes averaged; the 8-view soft-dropout variants use five scenes, while the full model uses all six. Best and second-best values are bold and underlined.

Figure 3: Qualitative comparison on MiP-NeRF 360 (24/8 views) and LLFF (3 views). Each non-GT panel reports the displayed-view PSNR and final Gaussian count.

$P$	$S$	$R$	$V$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
5	0	0	0	12.8370	0.4092	0.6619
5	0	0	1	14.7173	0.4575	0.6400
6	0	0	0	15.0757	0.4628	0.6396
6	1	1	0	14.6207	0.4508	0.6461
7	0	0	0	10.9079	0.4070	0.6758

Table 3: Controlled HashGrid input-encoding ablation on MiP-NeRF 360 kitchen with 24 training views.

P

S

R

, and

V

are the numbers of encoding levels assigned to position, scale, rotation, and view direction. Bold indicates the best result.

Figure 4: Uncertainty dynamics for the LLFF fern scene with three training views: raw-uncertainty distributions (left), the fraction with $\tilde{u}_{i}>0.8$ (centre), and the uncertainty–opacity response at 6k iterations (right).

Figure 5: Photometric residuals for 3DGS*, DropGS, and FSGS, and UGOD’s predicted uncertainty. Rows show MiP-NeRF 360 bonsai (24 views) and LLFF horns and fern (3 views).

Symbol	Description	Value
$\kappa$	modulation slope	$4.0$
$\tau_{g}$	modulation threshold	$0.8$
$\beta$	modulation floor (min. opacity multiplier)	$0.70$
$T_{g}$	modulation warm-up (iters)	$1200$
$\eta$	dropout scale	$0.08$
$\tau_{d}$	dropout threshold	$0.0$

Table 1: Default hyper-parameters for the uncertainty head, opacity modulation, detached soft dropout, and early stopping.

Scene	Method	Init	Metrics			Final Gaussians	Reduction
Scene	Method	Init	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Final Gaussians	vs 3DGS*
bicycle_8	3DGS*	3	11.01	0.173	0.592	1,460,812	-
	DropGS	3	12.24	0.208	0.597	1,000,235	1.5 $\times$ fewer
	FSGS (w/ depth)	3	11.05	0.202	0.626	142,066	10.3 $\times$ fewer
	Ours (full)	3	13.11	0.284	0.643	59,483	24.6 $\times$ fewer
bonsai_8	3DGS*	11	11.9641	0.3527	0.5958	565,053	-

Table 2: Per-scene results on MiP-NeRF 360 (8 views). The full model is shaded; best and second-best values are bold and underlined.

Scene	Method	Init	Metrics			Final Gaussians	Reduction
Scene	Method	Init	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Final Gaussians	vs 3DGS*
bicycle_24	3DGS*	21	15.3990	0.2815	0.5713	1,292,646	-
	DropGS	21	15.8850	0.3172	0.5894	937,115	1.4 $\times$ fewer
	FSGS (w/ depth)	21	15.8226	0.3482	0.6068	155,215	8.3 $\times$ fewer
	Ours (full)	21	16.8980	0.3602	0.5954	141,530	9.1 $\times$ fewer
bonsai_24	3DGS*	5764	22.5979	0.8081	0.2986	686,826	-

Table 3: Per-scene results on MiP-NeRF 360 (24 views). The full model is shaded; best and second-best values are bold and underlined.

Scene	Method	Init	Metrics			Final	Compression
		Points	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Gaussians	vs 3DGS*
Fern	3DGS*	1707	13.99	0.444	0.421	325,283	–
	DropGS	1707	13.69	0.448	0.425	245,406	1.3 $\times$ fewer
	FSGS (w/ depth)	1707	20.49	0.665	0.243	117,432	2.8 $\times$ fewer
	Ours (full)	1707	20.74	0.680	0.231	93,766	3.5 $\times$ fewer
Flower	3DGS*	1804	16.81	0.479	0.371	233,646	–

Table 4: Per-scene results on LLFF (3 views). The full model is shaded; best and second-best values are bold and underlined.

Scene	Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Final Gaussians	vs 3DGS*
	Ours (w/o soft dropout)	11.24	0.239	0.688	24,313	60.1 $\times$ fewer
	Ours (w/ soft dropout)	10.56	0.246	0.689	14,400	101.4 $\times$ fewer
bicycle	Ours (full)	13.11	0.284	0.643	59,483	24.6 $\times$ fewer
	Ours (w/o soft dropout)	13.9212	0.4432	0.5847	97,801	5.8 $\times$ fewer
	Ours (w/ soft dropout)	12.6963	0.4385	0.6108	52,688	10.7 $\times$ fewer
bonsai	Ours (full)	13.7776	0.4309	0.5643	111,546	5.1 $\times$ fewer

Table 5: Per-scene component ablation on MiP-NeRF 360 (8 views). Best and second-best values are bold and underlined.

Scene	Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Final Gaussians	vs 3DGS*
bicycle	Ours (w/o soft dropout)	16.6135	0.3429	0.6026	84,962	15.2 $\times$ fewer
	Ours (w/ soft dropout)	15.7267	0.3208	0.6139	76,494	16.9 $\times$ fewer
	Ours (full)	16.8980	0.3602	0.5954	141,530	9.1 $\times$ fewer
bonsai	Ours (w/o soft dropout)	22.1005	0.7864	0.3320	337,290	2.0 $\times$ fewer
	Ours (w/ soft dropout)	21.7632	0.7790	0.3339	344,817	2.0 $\times$ fewer
	Ours (full)	20.1803	0.7038	0.4269	154,205	4.5 $\times$ fewer

Table 6: Per-scene component ablation on MiP-NeRF 360 (24 views). Best and second-best values are bold and underlined.

Scene	Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Final Gaussians	vs 3DGS*
Fern	Ours	20.09	0.649	0.248	198,554	1.6 $\times$ fewer
	Ours (w/o soft dropout)	19.06	0.605	0.277	204,963	1.6 $\times$ fewer
	Ours (full)	20.74	0.680	0.231	93,766	3.5 $\times$ fewer
Flower	Ours	19.85	0.619	0.251	193,944	1.2 $\times$ fewer
	Ours (w/o soft dropout)	18.97	0.591	0.271	192,774	1.2 $\times$ fewer
	Ours (full)	19.96	0.616	0.264	93,361	2.5 $\times$ fewer

Table 7: Per-scene component ablation on LLFF (3 views). Best and second-best values are bold and underlined.

Explore similar work

From Uncertainty to Stability and Fidelity: Guiding Sparse-View 3D Gaussian Splatting with Fisher Information
Jun 18, 2026Junbao Zhou, Qingshan Xu, Yuan Zhou +7Novel View Synthesis 3D Gaussian
Nanyang Technological University, Singapore · University of Science and Technology of China, China · Zhejiang University, China +1
Rendering-Aware Bayesian 3D Gaussian Splatting with Native Uncertainty and Adaptive Complexity Control
Jul 6, 2026Gaoxiang Jia, Vikram Appia, Junzhou Huang +13D Gaussian Novel View Synthesis
Advanced Micro Devices, Inc. · Dept. of Computer Science and Engineering University of Texas at Arlington · Dept. of Mathematics & Division of Data Science University of Texas at Arlington
SparseGS: Sparse View Synthesis using 3D Gaussian Splatting
Nov 30, 2023Haolin Xiong, Sairisheek Muttukuru, Hanyuan Xiao +4Novel View Synthesis Synthesis
University of California, Los Angeles · USC Institute for Creative Technologies · University of Southern California

Jun 18, 2026, cs.CVJ/K move · Enter open · S save

From Uncertainty to Stability and Fidelity: Guiding Sparse-View 3D Gaussian Splatting with Fisher Information

Junbao Zhou, Qingshan Xu, Yuan Zhou, Xiaolong Shen+6

Nanyang Technological University, Singapore · University of Science and Technology of China, China · Zhejiang University, China+1

Novel View Synthesis 3D Gaussian Gaussian Splatting Geometric Priors Regularization Fisher Information

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) has emerged as a promising technique for novel view synthesis. However, 3DGS requires dense input views to achieve high-quality rendering. In sparse-view scenarios, 3DGS often prones to overfitting, resulting in noticeable artifacts and degraded rendering quality. Previous methods explore to address this issue by introducing additional priors (e.g. depth priors) or integrating regularization techniques (e.g. Dropout). However, these methods are often applied without principled guidance. In particular, prior-based augmentation typically samples novel viewpoints randomly, while Dropout-based regularization randomly removes Gaussians. The compounded randomness introduces uncertainty and instability, limiting the fidelity of novel view synthesis. In this paper, we propose a novel method for sparse-view 3DGS that incorporates Fisher Information to quantitatively guide the utilization of geometric priors and regularization. Specifically, our method comprises two key components: (1) Stereo augmentation with Fisher Information. By leveraging Fisher Information, we actively select most informative supporting views and use depth priors to curate reliable pseudo ground truths, which reduces randomness in augmentation and improves stability and rendering fidelity; (2) Uncertainty-aware regularization. We reduce the instability of Dropout-based regularization by using Fisher Information to quantitatively measure the uncertainty of each 3D Gaussian, and adaptively adjust the removal probability, leading to more stable and effective regularization. With these two components, our method effectively mitigates overfitting and improves the stability of optimization in sparse-view 3DGS, resulting in superior rendering fidelity. Extensive experiments show that our method achieves state-of-the-art performance in sparse-view novel view synthesis benchmarks.

Abstract

Explore similar work

From Uncertainty to Stability and Fidelity: Guiding Sparse-View 3D Gaussian Splatting with Fisher Information

Rendering-Aware Bayesian 3D Gaussian Splatting with Native Uncertainty and Adaptive Complexity Control

SparseGS: Sparse View Synthesis using 3D Gaussian Splatting