notesum.ai

Published at December 6

UniMLVG: Unified Framework for Multi-view Long Video Generation with Comprehensive Control Capabilities for Autonomous Driving

cs.CV

Released Date: December 6, 2024

Authors: Rui Chen¹, Zehuan Wu, Yichen Liu, Yuxin Guo, Jingcheng Ni, Haifeng Xia, Siyu Xia

Aff.: ¹School of Automation, Southeast University, China

Arxiv: http://arxiv.org/pdf/2412.04842v1

Refer to caption

Method	Multi-view	Video	Duration	FID $\downarrow$	FVD $\downarrow$	$\textbf{mAP}_{obj}$ $\uparrow$	$\textbf{mIoU}_{road}$ $\uparrow$	$\textbf{mIoU}_{vehicle}$ $\uparrow$
Oracle	-	-	-	-	-	35.56	73.67	31.86
DriveGAN [24]	$\times$	$\surd$	3s	73.4	502.3	-	-	-
DriveDreamer [49]	$\times$	$\surd$	4s	52.6	452.0	-	-	-
MagicDrive [12]	$\surd$	$\surd$	5s	19.1	218.1	12.30	61.05	27.01
Drive-WM [50]	$\surd$	$\surd$	20s	15.2	122.7	20.66	65.07	-
DriveDreamer-2 [62]	$\surd$	$\surd$	7s	11.2	55.7	-	-	-
DreamForge [30]	$\surd$	$\surd$	20s	16.0	224.8	13.80	-	-
DiVE [23]	$\surd$	$\surd$	20s	-	94.6	24.55	-	-
Ours	$\surd$	$\surd$	20s	8.8	60.1	22.50	70.81	29.12