notesum.ai

Published at December 9

Generative Densification: Learning to Densify Gaussians for High-Fidelity Generalizable 3D Reconstruction

cs.CV

cs.GR

Released Date: December 9, 2024

Authors: Seungtae Nam¹, Xiangyu Sun¹, Gyeongjin Kang¹, Younggeun Lee¹, Seungjun Oh¹, Eunbyung Park¹

Aff.: ¹Sungkyunkwan University

Arxiv: http://arxiv.org/pdf/2412.06234v1

[Uncaptioned image]

Method	#Param(M)	Gobjaverse [33]			GSO [10]			Co3D [25]
Method	#Param(M)	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
MVSNeRF [4]	0.52	14.48	0.896	0.185	15.21	0.912	0.154	12.94	0.841	0.241
MuRF [34]	15.7	14.05	0.877	0.301	12.89	0.885	0.279	11.60	0.815	0.393
LGM [30]	415	19.67	0.867	0.157	23.67	0.917	0.063	13.81	0.739	0.414
GS-LRM [42]	300	-	-	-	30.52	0.952	0.050	-	-	-
LaRa [5]	125	27.49	0.938	0.093	29.70	0.959	0.060	21.18	0.862	0.216
Ours-fast	134	28.23	0.943	0.084	30.62	0.964	0.062	21.67	0.864	0.209
Ours	134	28.58	0.945	0.080	31.06	0.966	0.058	21.72	0.865	0.209
Ours (w/ residual)	134	28.75	0.946	0.078	31.23	0.967	0.058	22.08	0.867	0.206