notesum.ai

Published at November 26

Buffer Anytime: Zero-Shot Video Depth and Normal from Image Priors

cs.CV

cs.AI

Released Date: November 26, 2024

Authors: Zhengfei Kuang¹, Tianyuan Zhang², Kai Zhang³, Hao Tan³, Sai Bi³, Yiwei Hu³, Zexiang Xu³, Milos Hasan³, Gordon Wetzstein¹, Fujun Luan³

Aff.: ¹Stanford University; ²Massachusetts Institute of Technology; ³Adobe Research

Arxiv: http://arxiv.org/abs/2411.17249v1

[Uncaptioned image]

Method	Time	ScanNet [11]				KITTI [21]				Bonn [36]
Method	Time	AbsRel $\downarrow$	$\delta_{1}$ $\uparrow$	OPW $\downarrow$		AbsRel $\downarrow$	$\delta_{1}$ $\uparrow$	OPW $\downarrow$		AbsRel $\downarrow$	$\delta_{1}$ $\uparrow$	OPW $\downarrow$
ChronoDepth [45]	106s	0.159	0.783	0.092		0.151^∗	0.797^∗	0.050		0.109^∗	0.886^∗	0.035
NVDS [53]	283s	0.187	0.677	0.143		0.253	0.588	0.089		0.210^∗	0.693^∗	0.068
DepthCrafter [30]	270s	0.125	0.848	0.082		0.110	0.881	0.111		0.075	0.971	0.029
MariGold [32]	475s	0.166	0.769	0.241		0.149	0.796	0.235		0.091	0.931	0.109
MariGold-E2E-FT [20]	72s	0.150	0.802	0.145		0.151	0.779	0.100		0.090	0.921	0.053
Depth Anything V2 [56]	31s	0.135	0.822	0.121		0.140	0.804	0.089		0.119^∗	0.875^∗	0.059
Ours (Depth Anything V2)	33s	0.123	0.853	0.076		0.119	0.865	0.038		0.102	0.925	0.028