EndoPrior-GS: Dynamic Endoscopic Reconstruction with a Joint Texture Prior

Jiaqi Huang; Shidong Wang; Tong Xin; Kabita Adhikari

cs.CVSep 29, 2026

EndoPrior-GS: Dynamic Endoscopic Reconstruction with a Joint Texture Prior

Authors: Jiaqi Huang, Shidong Wang, Tong Xin, Kabita Adhikari

Organizations: School of Engineering, Newcastle University, Newcastle upon Tyne NE1 7RU, UK · School of Computing, Newcastle University, Newcastle upon Tyne NE4 5TG, UK

Abstract

Dynamic endoscopic reconstruction is fundamental to robotic surgery and computer-assisted interventions. While 3D Gaussian Splatting (3DGS) realises real-time rendering, its application to deformable intraoperative environments remains constrained by spurious geometry and varying illuminations. To address these limitations, we introduce EndoPrior-GS, a novel pipeline that explicitly couples frame-extracted vision heuristics and estimated depth maps. EndoPrior-GS derives a joint texture prior from a tool-filtered valid tissue mask, a non-specular photometric filter, and anatomical structural salience, yielding a probability map that guides primitive initialisation and subsequent density control. The prior is further extended to the temporal domain through a texture-aware term that dynamically weighs pairwise primitive contributions during training. We conduct extensive experiments on benchmark datasets EndoNeRF and SCARED, and the obtained results show that our method EndoPrior-GS reduces Flow Error by 27.7% and 25.8% over the representative approaches while preserving competitive rendering quality and real-time rendering speed. Our project website is available at https://jiaqi-huang-77.github.io/EndoPrior-GS/.

Figures & tables

Figure 1 : Comparison of rendering results in challenging endoscopic regions. Panels (a)–(c) show extreme specular highlights, extreme dark regions, and tissue boundaries, respectively. Compared with EndoGaussian [ 13 ] , EndoPrior-GS (Ours) mitigates semi-transparent floaters, oversized Gaussians near tissue boundaries, and degenerate sheet-like Gaussians in low-texture regions.

Figure 2 : Efficiency and Stability. EndoPrior-GS (Ours) achieves lower Flow Error with fewer Gaussians.

Figure 3 : The proposed EndoPrior-GS framework.

Figure 4 : Joint Texture Prior. $E_{t}$ , $L_{t}^{\mathrm{sp}}$ , and $\rho_{t}=M_{t}\odot(1-L_{t}^{\mathrm{sp}})\odot E_{t}$ .

Figure 5 : Prior-Guided Densification. The densification trigger is modulated by the texture-prior score $\bar{s}_{i}$ .

Figure 6 : Temporal Regularisation.

Figure 7 : Qualitative comparison with representative endoscopic reconstruction methods. Results are shown on selected pulling and cutting sequences. Rows denote timestamps and columns denote methods or ground truth.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	RMSE $\downarrow$	Flow Error $\downarrow$	FPS $\uparrow$
EndoNeRF
EndoNeRF [ 23 ] 1	32.039	0.925	0.081	–	–	–
	27.519	0.933	0.128	2.522	5.421	0.14
EndoSurf [ 27 ] 2	34.992	0.954	0.113	1.000	–	–
	35.004	0.845	0.243	2.584	5.284	12.84
EndoGaussian [ 13 ] 5	37.849	0.963	0.054	–	–	–

Table 1 : Quantitative comparison on EndoNeRF [ 23 ] and SCARED [ 2 ] . Grey cells denote original-paper values, followed by our reproduced results under a unified protocol. Red superscripts indicate official GitHub links. Best and second-best results are highlighted in red and orange.

Budget	Method	$N_{\mathrm{GS}}$ $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	RMSE $\downarrow$	Flow Error $\downarrow$	FPS $\uparrow$
10K	EndoGaussian [ 13 ]	14.8K	36.81	0.953	0.084	2.97	4.46	307
	EndoPrior-GS	9.7K	36.87	0.951	0.095	2.47	2.82	310
20K	EndoGaussian [ 13 ]	22.5K	37.14	0.957	0.066	2.80	4.11	295
	EndoPrior-GS	21.6K	38.07	0.967	0.071	2.26	2.74	305
30K	EndoGaussian [ 13 ]	30.7K	37.32	0.959	0.058	2.48	3.79	280
	EndoPrior-GS	26.9K	38.12	0.973	0.067	2.27	3.07	300

Table 2: Efficiency and stability comparison under matched initialisation budgets. The budget denotes the nominal initialisation setting, while

N_{\mathrm{GS}}

is the Gaussian count after training.

Seq.	EndoG. [ 13 ]	EndoPrior-GS	Reduction
Pulling	0.4459	0.3786	15.1%
Cutting	0.1166	0.1071	8.1%

Table 3 : Average rendered-depth temporal instability diagnostic.

Figure 8 : Comparison of rendered-depth temporal instability diagnostic maps. Heat-maps visualise representative pixel-wise adjacent-frame depth variation.

Variant	$N_{\mathrm{GS}}\downarrow$	$N_{\mathrm{low}\text{-}prior}$	$N_{\mathrm{low}\text{-}\alpha}\downarrow$	$N_{\mathrm{large}\text{-}scale}\downarrow$	Flow Error $\downarrow$
Baseline	22.5k	19.9k	262.0	5.0	4.112
w/o Texture-aware temp. reg.	21.5k	19.9k	110.3	0.7	3.152
EndoPrior-GS (Ours)	21.6k	19.9k	105.0	0.0	2.742

Table 4 : Quantitative primitive count diagnostic analysis. Counts use fixed post-hoc thresholds after training. The baseline denotes EndoGaussian [ 13 ] under 20K initialisation budgets.

Figure 9 : Qualitative primitive distribution analysis. We show input RGB, joint texture prior $\rho_{t}$ , and projected Gaussian density, with diagnostic markers computed using the thresholds in Table 4 .

Variant	Pair. w.	Disp. c.	Dist. c.	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	RMSE $\downarrow$	Flow Error $\downarrow$
Baseline	$\times$	$\times$	$\times$	37.147	0.957	0.066	2.804	4.112
w/o Pairwise weighting	$\times$	$\checkmark$	$\checkmark$	37.017	0.959	0.068	2.785	3.569
w/o Distance control	$\checkmark$	$\checkmark$	$\times$	37.104	0.959	0.070	2.797	3.144
w/ Temporal reg.	$\checkmark$	$\checkmark$	$\checkmark$	37.129	0.961	0.069	2.575	3.003

Table 5: Ablation of texture-aware temporal regularisation. Pair. w.: pairwise affinity weighting; Disp. c.: primitive displacement control; Dist. c.: pairwise distance control.

Variant	Pri. init.	Pri. dens.	Temp. reg.	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	RMSE $\downarrow$	Flow Error $\downarrow$
Prior init.	$\checkmark$	$\times$	$\times$	37.080	0.959	0.060	2.727	3.343
+ Prior dens.	$\checkmark$	$\checkmark$	$\times$	37.165	0.961	0.065	2.479	3.152
Full model	$\checkmark$	$\checkmark$	$\checkmark$	38.041	0.967	0.071	2.260	2.701

Table 6: Ablation study of the components. All variants are evaluated under the matched 20K budget.

Explore similar work

ExtraGS: Enhancing Endoscopic View Extrapolation via Diffusion-Guided 3D Gaussian Splatting
Jul 14, 2026Cheng-Tai Hsieh, Jiwei Shan, Han Fang +6Novel View Synthesis Colonoscopy
School of Automation and Intelligent Sensing, Shanghai Jiao Tong University, and the Shanghai Key Laboratory of Navigation and Location-Based Services, Shanghai, China · Department of Mechanical and Automation Engineering and the T Stone Robotics Institute, The Chinese University of Hong Kong, Hong Kong · Global College, Shanghai Jiao Tong University, Shanghai, China +1
Endo-NeRF++: Uncertainty-Aware Neural Rendering with Multi-Resolution Hash Encoding for Dynamic Surgical Scene Reconstruction
Jul 30, 2026Gousia Habib, Laura RuotsalainenNeural Rendering Scene Understanding
Department of Computer Science, University of Helsinki, Finland
EvoGS: Modeling Deformation Evolution for Dynamic Gaussian Splatting
Sep 1, 2026Wei Dong, Shahram Shirani, Jun Chen +1Deformable Cnn Scene Reconstruction
The Department of Electrical & Computer Engineering, McMaster University

Jul 14, 2026, cs.CVJ/K move · Enter open · S save

ExtraGS: Enhancing Endoscopic View Extrapolation via Diffusion-Guided 3D Gaussian Splatting

Cheng-Tai Hsieh, Jiwei Shan, Han Fang, Jianshu Hu+5

School of Automation and Intelligent Sensing, Shanghai Jiao Tong University, and the Shanghai Key Laboratory of Navigation and Location-Based Services, Shanghai, China · Department of Mechanical and Automation Engineering and the T Stone Robotics Institute, The Chinese University of Hong Kong, Hong Kong · Global College, Shanghai Jiao Tong University, Shanghai, China+1

Novel View Synthesis Colonoscopy Neural Radiance Fields Extrapolation

Open paper AskPDF arXiv

Robot-assisted minimally invasive surgery (MIS) critically depends on reliable endoscopic perception for navigation and safety. However, conventional endoscopes provide only a limited field of view, leaving large portions of the surrounding anatomy unobserved. Recent neural rendering approaches, such as Neural Radiance Fields and 3D Gaussian Splatting, enable novel view synthesis from endoscopic videos, but their reliance on sparse observations often leads to severe artifacts when extrapolating beyond the training trajectory. In this work, we propose ExtraGS, a framework for enhancing endoscopic view extrapolation through diffusion-guided 3D Gaussian Splatting. Starting from an initial reconstruction, we introduce an uncertainty-guided virtual camera sampling strategy to actively explore blind spots and maximize information gain. The rendered views from these sampled locations are refined using a diffusion model to recover plausible anatomical structures, producing pseudo-observations that guide further optimization. To prevent the generated content from degrading reliable regions, we adopt a confidence-weighted fine-tuning strategy when incorporating these pseudo-observations. Extensive experiments on multiple public endoscopic datasets demonstrate that ExtraGS significantly reduces extrapolation artifacts and achieves state-of-the-art performance in endoscopic novel view synthesis.

Abstract

Explore similar work

ExtraGS: Enhancing Endoscopic View Extrapolation via Diffusion-Guided 3D Gaussian Splatting

Endo-NeRF++: Uncertainty-Aware Neural Rendering with Multi-Resolution Hash Encoding for Dynamic Surgical Scene Reconstruction

EvoGS: Modeling Deformation Evolution for Dynamic Gaussian Splatting