GOFish: Full Field-of-View Gaussian Splatting and Mesh Extraction

Allen Tu; Meher Gitika Karumuri; Shifan Zhu; Hailing Wang; Yen-Liang Lin; David Held; Matthias Zwicker; Tom Goldstein

cs.CVOct 27, 2026

GOFish: Full Field-of-View Gaussian Splatting and Mesh Extraction

Authors: Allen Tu, Meher Gitika Karumuri, Shifan Zhu, Hailing Wang, Yen-Liang Lin, David Held, Matthias Zwicker, Tom Goldstein

Organizations: Amazon · University of Maryland, College Park · University of Massachusetts Amherst · Northeastern University ·

Abstract

Robots, autonomous vehicles, and consumer 360° cameras use fisheye lenses for wide fields of view (FoVs). Simulation-ready digital twins require photorealistic rendering across the full FoV and an aligned collision mesh. However, prior fisheye Gaussian Splatting methods only rasterize up to 180° and do not produce meshes, while Gaussian surface reconstruction methods assume pinhole training images. We present GOFish, the first 3D Gaussian Splatting surface reconstruction method to train directly on fisheye images, rasterize FoVs beyond 180°, and extract an aligned mesh from the same scene. Calibrated densification recovers the surface across the entire FoV, while Calibrated Mip Filtering and Wide-Angle Appearance Modeling keep pixel blur and lens falloff out of the scene and mesh. Across the 200° FIORD and 205° FullCircle benchmarks, GOFish improves PSNR over the best prior method by 2.28 and 1.97 dB, respectively. It also outperforms fisheye rendering baselines within their own FoVs and achieves the highest mesh F-score on FIORD and the main ScanNet++ set. Together, the Gaussian scene and mesh form a LiDAR-free, simulation-ready digital twin that renders in real time, ray traces the full FoV, and provides collision geometry.

Figures & tables

		FIORD [ 8 ]						FullCircle [ 5 ]						ScanNet++ [ 34 ]
		Full ( $200^{\circ}$ )			GOFish @ FoV			Full ( $205^{\circ}$ )			GOFish @ FoV			Full ( $127^{\circ}$ )
Method	FoV	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\Delta$ PSNR $\uparrow$	$\Delta$ SSIM $\uparrow$	$\Delta$ LPIPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\Delta$ PSNR $\uparrow$	$\Delta$ SSIM $\uparrow$	$\Delta$ LPIPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Fisheye-GS ∗ [ 22 ]	$145^{\circ}$	17.78	0.685	0.395	+5.06	+0.119	+0.167	14.01	0.630	0.495	+14.05	+0.263	+0.397	26.28	0.895	0.256
3DGUT [ 31 ]	$160^{\circ}$	19.86	0.740	0.316	+3.95	+0.057	+0.046	21.58	0.842	0.262	+1.62	+0.013	+0.030	28.93	0.913	0.214
3DGUT + PPISP [ 31 , 4 ]	$160^{\circ}$	20.18	0.743	0.312	+4.04	+0.058	+0.048	21.45	0.841	0.262	+1.88	+0.014	+0.030	28.93	0.913	0.213
3DGEER ∗ [ 14 ]	$178^{\circ}$	12.79	0.679	0.328	+2.03	+0.026	+0.027	12.65	0.715	0.336	+4.25	+0.040	+0.036	28.59	0.914	0.201

Table 1: Quantitative comparison against wide-angle rendering baselines. The Ag best , Ag second best , and Ag third best entries are highlighted. GOFish’s gain at each baseline’s FoV is Ag highlighted .

\ast

Appendix B.3 details baseline adaptations for wide-angle rendering.

Panel of Figure 2 — Figure 2: 3DGUT failures over a full field of view. (a) Its half-angle limit cuts rays beyond $80^{\circ}$ from the axis ( orange ), and its positive-depth test fails beyond $90^{\circ}$ ( red ). (b) Tiles come from the projected sigma-point extent ( blue ) but alpha from the exact ray response ( red ), so the Gaussian is skipped in the hatched tiles. (Bottom) 3DGUT + PPISP leaves a cutoff ring and tile-shaped smears, while GOFish renders the full FoV cleanly.

Panel of Figure 3 — Figure 3: Calibrated Mip Filtering. The finest training view’s Jacobian sets the 3D filter $+\epsilon_{i}^{2}\boldsymbol{I}_{3}$ in the stored Gaussian covariance $\boldsymbol{\Sigma}_{i}$ ( blue ). Rendering view $v$ adds the footprint $+\gamma\boldsymbol{J}^{+}\boldsymbol{J}^{+\top}$ and rescales opacity by $m$ ( orange ), while meshing uses $\boldsymbol{\Sigma}_{i}$ directly.

		FIORD [ 8 ]									ScanNet++ [ 34 ]
		Full ( $200^{\circ}$ )			Common ( $90^{\circ}$ )			Geometry			Full ( $127^{\circ}$ )			Geometry
Method	FoV	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$P\uparrow$	$R\uparrow$	$F\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$P\uparrow$	$R\uparrow$	$F\uparrow$
GOF [ 37 ]	Rectified	17.03	0.676	0.353	28.20	0.883	0.278	0.168	0.248	0.198	27.54	0.904	0.212	0.579	0.629	0.601
PGSR [ 2 ]	Rectified	16.81	0.670	0.363	28.44	0.892	0.273	0.225	0.293	0.239	27.39	0.898	0.241	0.698	0.677	0.687
SOF [ 26 ]	Rectified	16.62	0.660	0.363	26.15	0.852	0.311	0.163	0.240	0.192	27.10	0.898	0.218	0.620	0.667	0.641
MiLo [ 7 ]	Rectified	17.05	0.680	0.368	28.21	0.882	0.297	0.180	0.261	0.210	27.61	0.903	0.234	0.651	0.701	0.675

Table 2: Quantitative comparison against surface reconstruction baselines. The Ag best , Ag second best , and Ag third best entries are highlighted. Full scores each calibrated FoV and Common the shared rectified images; Section 5 defines both.

Panel of Figure 4 — Figure 4: Qualitative rendering results on FIORD [ 8 ] and FullCircle [ 5 ] . Held-out views span the full FoV; metrics are per-scene.

Panel of Figure 5 — Figure 5: Qualitative mesh results on FIORD [ 8 ] and FullCircle [ 5 ] . GOFish uses the Geometry Objective. Metrics are per-scene.

Panel of Figure 6 — Figure 6: Ray tracing and mesh-based physics simulation. 3DGUT ray traces its $160^{\circ}$ FoV using 3DGRT [ 24 ] and does not produce a mesh, so simulated objects have no surface to collide with. GOFish covers the full $200^{\circ}$ FoV and supports collisions.

Figure 7: Rendering gain by ray angle on FIORD [ 8 ] ( cyan ) and FullCircle [ 5 ] ( magenta ). PSNR gain of GOFish over 3DGUT + PPISP [ 31 , 4 ] (solid) and over GOFish without Wide-Angle Appearance (dashed). As in Fig. 2 , 3DGUT + PPISP cuts rays beyond

80^{\circ}

( orange ) and fails beyond

90^{\circ}

( red ).

Variant	FoV	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	F $\uparrow$	Train $\downarrow$	FPS $\uparrow$
GOF [ 37 ]	90°	17.03	0.676	0.353	0.198	104m	11.0
3DGRT + PPISP [ 24 , 4 ]	Full	18.42	0.712	0.313	0.159	108m	22.3
3DGUT + SOF [ 31 , 26 ]	160°	18.46	0.671	0.374	0.168	111m	17.9
3DGUT + PPISP [ 31 , 4 ]	160°	20.18	0.743	0.312	0.141	29m	84.3
+ Full-FoV Rasterization	Full	21.21	0.798	0.253	0.139	162m	91.1
+ Calibrated Mip Filtering	Full	21.19	0.804	0.239	0.240	103m	81.1

Table 3: Component analysis on FIORD [ 8 ] . Components are added cumulatively. + Calibrated Mip Filtering includes calibrated densification, and + Wide-Angle Appearance completes base GOFish. Appendix B.2 details the configurations.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 8 — Figure 8: FIORD capture and calibration diagnostics [ 8 ] . (Top) SfM reconstructions are overlaid on LiDAR ( gray ). Points within 10 cm of LiDAR are green and others orange . Night_Out has no LiDAR. Cameras are black ; red marks those more isolated than any retained camera, including one off-panel (arrow). (a) Full-FoV 95th-percentile training luma (left) and a rim crop (right) for Building_In camera 2. Curves mark total field angles of $167^{\circ}$ ( blue ) and $200^{\circ}$ ( orange ), while magenta marks the dark band. (b) Held-out ground truth (left) and GOFish rendering (right) at the rim, with the calibrated edge in white.

	ScanNet++ (3DGUT set) [ 34 , 31 ]									ScanNet++ (CoMe set) [ 34 , 27 ]
	Full ( $127^{\circ}$ )			Common			Geometry			Full ( $127^{\circ}$ )			Common			Geometry
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$P\uparrow$	$R\uparrow$	$F\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$P\uparrow$	$R\uparrow$	$F\uparrow$
GOF [ 37 ]	27.54	0.904	0.212	30.01	0.920	0.208	0.579	0.629	0.601	26.10	0.874	0.234	27.34	0.889	0.233	0.578	0.693	0.630
PGSR [ 2 ]	27.39	0.898	0.241	29.66	0.916	0.229	0.698	0.677	0.687	25.69	0.870	0.258	26.94	0.887	0.249	0.694	0.728	0.710
SOF [ 26 ]	27.10	0.898	0.218	28.83	0.914	0.214	0.620	0.667	0.641	26.29	0.876	0.221	27.20	0.889	0.225	0.603	0.724	0.657
MiLo [ 7 ]	27.61	0.903	0.234	29.75	0.916	0.231	0.651	0.701	0.675	26.35	0.877	0.247	27.50	0.891	0.247	0.652	0.726	0.687

Appendix

Table 4: Quantitative comparison on both ScanNet++ scene sets [ 34 ] . Left uses the six scenes of 3DGUT and Fisheye-GS [ 31 , 22 ] ; right uses CoMe’s six [ 27 ] . Full is the native

127^{\circ}

image and Common the official undistortion. GOFish is mapped to Common with the official map and a

2{\times}

box downsample; geometry uses CoMe’s 5 cm protocol. The Ag best , Ag second best , and Ag third best entries are highlighted.

	Rendering			Geometry
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$P\uparrow$	$R\uparrow$	$F\uparrow$
Fisheye-GS ∗ [ 22 ]	24.71	0.838	0.236
3DGUT [ 31 ]	25.11	0.852	0.221
3DGUT + PPISP [ 31 , 4 ]	26.51	0.868	0.199
3DGEER ∗ [ 14 ]	24.42	0.837	0.235
Self-Calibrating GS ∗ [ 3 ]	22.66	0.759	0.332

Appendix

Table 5: Quantitative comparison on Tanks & Temples [ 18 ] . The Ag best , Ag second best , and Ag third best entries are highlighted.

\ast

Appendix B.3 details baseline adaptations.

FIORD [ 8 ]

ScanNet++ [ 34 ]

Tanks & Temples [ 18 ]

Method

Train

\downarrow

Mesh

\downarrow

Total

\downarrow

FPS

\uparrow

Gaussians

Triangles

Train

\downarrow

Mesh

\downarrow

Total

\downarrow

FPS

\uparrow

Gaussians

Triangles

Train

\downarrow

Mesh

\downarrow

Total

\downarrow

FPS

\uparrow

Gaussians

Triangles

3DGS [ 17 ]

23m

124

0.95M

16m

168

0.97M

22m

107

1.31M

Fisheye-GS ∗ [ 22 ]

42m

135

0.60M

119m

74.0

0.96M

21m

115

1.49M

3DGUT [ 31 ]

28m

92.1

1.03M

10m

238

0.37M

24m

105

1.00M

3DGUT + PPISP [ 31 , 4 ]

29m

84.3

1.03M

11m

209

0.35M

25m

96.5

1.00M

3DGEER ∗ [ 14 ]

51m

46.7

1.90M

67m

112

1.41M

71m

97.2

4.27M

Appendix

Table 6: Efficiency and representation size. Dataset averages follow Appendix B.2 . Training, meshing, and FPS are measured on an NVIDIA A10G GPU. GOFish FPS uses the real-time renderer detailed in Appendix A.6 .

\star

denotes an NVIDIA A100 GPU run detailed in Appendix B.3 .

Panel of Figure 9 — Figure 9: Representative surface-reconstruction failure cases. Panels (a)–(c) compare ground truth (left) with GOFish meshes (right). Panel (d) shows GOFish’s rendering (left) and mesh (right).

Dataset	Abbreviation	Scene
FIORD [ 8 ]	Kit.	Kitchen_In
	Cor.	Corridor_Out
	Road	Road_Out
	Bld.	Building_In
	Meet.	MeetingRoom_In
	Bri.	Bridge_Out

Appendix

Scene abbreviations for Tables 7 – 10 .

			FIORD [ 8 ]														FullCircle [ 5 ]
			Full ( $200^{\circ}$ )							GOFish @ FoV							Full ( $205^{\circ}$ )									GOFish @ FoV
	Method	FoV	Kit.	Cor.	Road	Bld.	Meet.	Bri.	Hall	Kit.	Cor.	Road	Bld.	Meet.	Bri.	Hall	Dark	F1	F2	Lab	Lou.	Per.	R1	R2	R3	Dark	F1	F2	Lab	Lou.	Per.	R1	R2	R3
PSNR $\uparrow$	Fisheye-GS ∗ [ 22 ]	$145^{\circ}$	18.72	16.22	17.19	18.18	16.28	17.47	20.42	+5.66	+2.93	+2.59	+6.56	+9.65	+3.43	+4.63	13.44	13.87	14.53	14.38	12.80	13.25	14.97	14.59	14.28	+15.05	+13.98	+14.05	+14.29	+8.24	+15.82	+15.51	+15.64	+13.88
	3DGUT [ 31 ]	$160^{\circ}$	20.20	17.62	19.49	19.03	19.62	24.25	18.84	+3.99	+4.01	+2.48	+6.36	+5.27	$-$ 1.51	+7.04	22.17	21.06	20.44	20.32	20.14	22.55	22.30	23.21	22.06	+1.40	+2.69	+1.15	+1.71	$-$ 1.46	+2.63	+2.13	+2.68	+1.63
	3DGUT + PPISP [ 31 , 4 ]	$160^{\circ}$	20.34	18.78	19.70	19.68	19.48	24.38	18.93	+4.36	+3.26	+2.14	+6.51	+6.12	$-$ 1.17	+7.09	22.03	20.99	20.58	20.31	20.00	22.06	22.09	23.07	21.93	+1.74	+2.53	+1.44	+1.82	$-$ 1.42	+3.13	+2.84	+3.00	+1.82
	3DGEER ∗ [ 14 ]	$178^{\circ}$	14.19	10.39	11.52	12.29	13.50	10.06	17.55	$-$ 0.39	+4.80	+3.55	+1.53	+0.46	+4.85	$-$ 0.60	13.45	12.96	10.02	10.36	13.07	14.75	13.57	13.47	12.19	+4.66	+3.70	+5.22	+5.34	+1.48	+4.35	+4.86	+4.87	+3.72

Appendix

Table 7: Per-scene rendering results on FIORD [ 8 ] and FullCircle [ 5 ] for fisheye baselines. The Ag best , Ag second best , and Ag third best entries are highlighted. GOFish’s gain at each baseline’s FoV is Ag highlighted . Protocols follow Appendix B.2 .

			FIORD [ 8 ]														FullCircle [ 5 ]
			Full ( $200^{\circ}$ )							Common ( $90^{\circ}$ )							Full ( $205^{\circ}$ )
	Method	FoV	Kit.	Cor.	Road	Bld.	Meet.	Bri.	Hall	Kit.	Cor.	Road	Bld.	Meet.	Bri.	Hall	Dark	F1	F2	Lab	Lou.	Per.	R1	R2	R3
PSNR $\uparrow$	GOF [ 37 ]	Rectified	18.02	17.42	14.92	17.64	19.35	12.86	19.03	30.96	24.15	22.44	28.19	31.88	30.60	29.20	21.48	22.62	19.94	22.81	18.67	22.97	22.36	22.95	22.77
	PGSR [ 2 ]	Rectified	17.82	17.12	14.56	18.36	18.81	12.59	18.43	30.37	25.14	24.34	29.31	31.00	29.34	29.54	21.08	23.35	20.04	23.30	18.31	23.08	22.46	23.10	23.59
	SOF [ 26 ]	Rectified	18.08	16.92	13.26	17.00	19.30	12.73	19.05	30.07	21.53	17.87	26.00	30.57	28.65	28.32	21.13	22.41	19.55	22.72	19.62	22.03	22.02	22.67	23.63
	MiLo [ 7 ]	Rectified	17.57	17.87	14.94	18.21	18.86	12.26	19.67	30.02	25.18	24.35	29.46	30.25	29.19	29.01	21.30	23.10	19.79	22.89	20.82	23.17	22.26	23.35	22.70

Appendix

Table 8: Per-scene rendering results on FIORD [ 8 ] and FullCircle [ 5 ] for surface baselines. The Ag best , Ag second best , and Ag third best entries are highlighted. Protocols follow Appendix B.2 .

			ScanNet++ [ 34 ]												Tanks & Temples [ 18 ]
			Full ( $127^{\circ}$ )						Common						Pinhole
	Method	FoV	0a5c	8d56	bb87	d415	e8ea	fe17	0a5c	8d56	bb87	d415	e8ea	fe17	Barn	Cat.	Court.	Ign.	Meet.	Truck
PSNR $\uparrow$	Fisheye-GS ∗ [ 22 ]	$145^{\circ}$	24.50	24.17	29.21	26.13	29.28	24.40							28.30	23.75	22.58	22.35	25.90	25.41
	3DGUT [ 31 ]	$160^{\circ}$	29.54	26.09	30.97	28.21	33.10	25.66							29.10	23.65	22.84	22.82	26.30	25.96
	3DGUT + PPISP [ 31 , 4 ]	$160^{\circ}$	29.84	25.94	30.99	28.11	33.16	25.54							30.18	26.00	24.02	25.01	27.06	26.79
	3DGEER ∗ [ 14 ]	$178^{\circ}$	27.50	26.29	31.22	28.37	32.16	26.01							28.63	23.50	22.91	21.84	25.23	24.39

Appendix

Table 9: Per-scene rendering results on ScanNet++ [ 34 ] and Tanks & Temples [ 18 ] . The Ag best , Ag second best , and Ag third best entries are highlighted. Protocols follow Appendix B.2 .

		FIORD [ 8 ]						ScanNet++ [ 34 ]						Tanks & Temples [ 18 ]
	Method	Kit.	Road	Bld.	Meet.	Bri.	Hall	0a5c	8d56	bb87	d415	e8ea	fe17	Barn	Cat.	Court.	Ign.	Meet.	Truck
Precision $\uparrow$	GOF [ 37 ]	0.488	0.016	0.108	0.292	0.031	0.074	0.539	0.546	0.615	0.647	0.562	0.566	0.522	0.402	0.432	0.726	0.400	0.585
	PGSR [ 2 ]	0.647	0.025	0.176	0.346	0.066	0.090	0.702	0.643	0.696	0.786	0.648	0.715	0.521	0.337	0.190	0.831	0.368	0.611
	SOF [ 26 ]	0.473	0.006	0.103	0.307	0.031	0.060	0.561	0.555	0.702	0.675	0.608	0.622	0.586	0.393	0.506	0.682	0.399	0.553
	MiLo [ 7 ]	0.532	0.012	0.137	0.307	0.027	0.065	0.705	0.590	0.619	0.706	0.644	0.646	0.551	0.368	0.413	0.774	0.426	0.602
	CoMe [ 27 ]	0.508	0.016	0.148	0.352	0.039	0.077	0.639	0.562	0.716	0.667	0.680	0.648	0.558	0.446	0.543	0.801	0.491	0.604

Appendix

Table 10: Per-scene geometry results. The Ag best , Ag second best , and Ag third best entries are highlighted. Protocols follow Appendix B.2 .

Panel of Figure 10 — Figure 10: Additional qualitative rendering results on FIORD [ 8 ] . Metrics are per-scene.

Panel of Figure 11 — Figure 11: Additional qualitative rendering and mesh results on FIORD [ 8 ] . Metrics are per-scene.

Panel of Figure 12 — Figure 12: Additional qualitative rendering results on FullCircle [ 5 ] . Metrics are per-scene.

Panel of Figure 13 — Figure 13: Qualitative rendering results on ScanNet++ [ 34 ] . Metrics are per-scene.

Panel of Figure 14 — Figure 14: Qualitative rendering and mesh results on ScanNet++ [ 34 ] . Metrics are per-scene.

Panel of Figure 15 — Figure 15: Qualitative rendering and mesh results on Tanks & Temples [ 18 ] . Metrics are per-scene.

3D Reconstruction

3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.

See all 957 papers

DirectFisheye-GS: Enabling Native Fisheye Input in Gaussian Splatting with Cross-View Joint Optimization
Apr 1, 2026Zhengxian Yang, Fei Xie, Xutao Xue +53D Gaussian Splatting 3D Reconstruction
BNRist, Tsinghua University · Shanghai AI Lab · Beihang University +1
Spherical-GOF: Geometry-Aware Panoramic Gaussian Opacity Fields for 3D Scene Reconstruction
Mar 9, 2026Zhe Yang, Guoqiang Zhao, Sheng Wu +23D Gaussian Splatting Panoramic 3D Reconstruction
School of Artificial Intelligence and Robotics, Hunan University, China · National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University, China
Fisheye3R: Adapting Unified 3D Feed-Forward Foundation Models to Fisheye Lenses
Mar 30, 2026Ruxiao Duan, Erin Hong, Dongxu Zhao +33D Foundation Models Domain Adaptation
Yale University · Google

Apr 1, 2026J/K move · Enter open · S save

DirectFisheye-GS: Enabling Native Fisheye Input in Gaussian Splatting with Cross-View Joint Optimization

Zhengxian Yang, Fei Xie, Xutao Xue, Rui Zhang+4

BNRist, Tsinghua University · Shanghai AI Lab · Beihang University+1

3D Gaussian Splatting 3D Reconstruction Multi-View 3D Reconstruction

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) has enabled efficient 3D scene reconstruction from everyday images with real-time, high-fidelity rendering, greatly advancing VR/AR applications. Fisheye cameras, with their wider field of view (FOV), promise high-quality reconstructions from fewer inputs and have recently attracted much attention. However, since 3DGS relies on rasterization, most subsequent works involving fisheye camera inputs first undistort images before training, which introduces two problems: 1) Black borders at image edges cause information loss and negate the fisheye's large FOV advantage; 2) Undistortion's stretch-and-interpolate resampling spreads each pixel's value over a larger area, diluting detail density -- causes 3DGS overfitting these low-frequency zones, producing blur and floating artifacts. In this work, we integrate fisheye camera model into the original 3DGS framework, enabling native fisheye image input for training without preprocessing. Despite correct modeling, we observed that the reconstructed scenes still exhibit floaters at image edges: Distortion increases toward the periphery, and 3DGS's original per-iteration random-selecting-view optimization ignores the cross-view correlations of a Gaussian, leading to extreme shapes (e.g., oversized or elongated) that degrade reconstruction quality. To address this, we introduce a feature-overlap-driven cross-view joint optimization strategy that establishes consistent geometric and photometric constraints across views-a technique equally applicable to existing pinhole-camera-based pipelines. Our DirectFisheye-GS matches or surpasses state-of-the-art performance on public datasets. Project Page: https://yzxqh.github.io/DirectFisheye-GS/ .

Abstract

3D Reconstruction

Explore similar work

DirectFisheye-GS: Enabling Native Fisheye Input in Gaussian Splatting with Cross-View Joint Optimization

Spherical-GOF: Geometry-Aware Panoramic Gaussian Opacity Fields for 3D Scene Reconstruction

Fisheye3R: Adapting Unified 3D Feed-Forward Foundation Models to Fisheye Lenses