notesum.ai

Published at December 10

ReCap: Better Gaussian Relighting with Cross-Environment Captures

cs.CV

Released Date: December 10, 2024

Authors: Jingzhi Li¹, Zongwei Wu¹, Eduard Zamfir¹, Radu Timofte¹

Aff.: ¹University of Wurzburg

Arxiv: http://arxiv.org/pdf/2412.07534v1

[Uncaptioned image]

	Unseen Envs							Training Envs		NVS
	night	snow	sunset	interior	fireplace	forest	average	courtyard	bridge	bridge*
PSNR $\uparrow$
GShader [18]	22.33	18.14	19.26	21.87	21.03	23.04	20.95	20.68	26.04	32.92
3DGS-DR [44]	20.88	18.92	23.53	21.00	19.52	25.19	21.51	20.58	33.19	35.50
GS-IR [25]	24.87	21.95	21.15	20.18	24.27	20.47	22.15	20.02	20.72	29.95
R3DG [15]	23.20	22.05	22.93	19.65	22.62	22.26	22.12	20.22	22.25	30.44
ReCap (ours)	26.68	24.56	26.56	24.34	25.51	25.45	25.52	25.51	26.97	32.03
SSIM $\uparrow$
GShader [18]	0.899	0.889	0.900	0.907	0.869	0.907	0.895	0.893	0.950	0.969
3DGS-DR [44]	0.892	0.897	0.931	0.909	0.858	0.910	0.900	0.894	0.971	0.978
GS-IR [25]	0.900	0.896	0.892	0.878	0.876	0.863	0.884	0.871	0.879	0.938
R3DG [15]	0.889	0.914	0.920	0.865	0.867	0.879	0.889	0.871	0.879	0.959
ReCap (ours)	0.932	0.938	0.951	0.929	0.903	0.927	0.930	0.942	0.951	0.965
LPIPS $\downarrow$
GShader [18]	0.086	0.115	0.095	0.091	0.106	0.089	0.102	0.097	0.061	0.043
3DGS-DR [44]	0.077	0.089	0.070	0.086	0.101	0.081	0.084	0.095	0.040	0.033
GS-IR [25]	0.104	0.102	0.105	0.112	0.112	0.114	0.108	0.122	0.113	0.071
R3DG [15]	0.104	0.088	0.081	0.126	0.113	0.103	0.108	0.113	0.113	0.053
ReCap (ours)	0.062	0.070	0.057	0.076	0.076	0.071	0.069	0.065	0.058	0.047