notesum.ai

Published at November 22

Dynamics-Aware Gaussian Splatting Streaming Towards Fast On-the-Fly Training for 4D Reconstruction

cs.CV

cs.AI

Released Date: November 22, 2024

Authors: Zhening Liu¹, Yingdong Hu¹, Xinjie Zhang¹, Jiawei Shao², Zehong Lin¹, Jun Zhang¹

Aff.: ¹Hong Kong University of Science and Technology; ²Institute of Artificial Intelligence (TeleAI), China Telecom

Arxiv: http://arxiv.org/abs/2411.14847v1

Refer to caption

Category	Method	PSNR $\uparrow$	DSSIM $\downarrow$	Train	Render $\uparrow$	Streamable
Category	Method	(dB)	DSSIM $\downarrow$	Time $\downarrow$	(FPS)	Streamable
Static	Plenoxels [13]	30.77	-	$>$ 1000 s	8.3	$\checkmark$
Static	I-NGP [34]	28.62	-	79 s	2.9	$\checkmark$
Offline	DyNeRF^† [26]	29.58	-	$>$ 1000 s	0.02	$\times$
	NeRFPlayer [37]	30.69	0.0340	75 s	0.05	$\checkmark$
	HexPlane [5]	31.70	-	140 s	0.21	$\times$
	K-Planes [14]	31.63	-	48 s	0.15	$\times$
	HyperReel [2]	31.10	0.0360	115 s	2	$\times$
	MixVoxels [42]	30.80	-	16 s	16.7	$\times$
	4DGS-Wu [47]	31.15	0.0331	10.85 s	30	$\times$
	4DGS-Yang [52]	32.01	0.0290	76 s	114	$\times$
	E-D3DGS [3]	31.31	0.0259	22.4 s	74.5	$\times$
	STG^‡ [27]	32.05	0.0261	60 s	140	$\times$
Online	StreamRF [25]	30.68	-	15 s	8.3	$\checkmark$
	3DGStream [38]	31.67	-	12 s	215	$\checkmark$
	DASS (Ours)	31.99	0.0285	9.62 s	189.2	$\checkmark$