notesum.ai

Published at October 30

cs.CV

cs.AI

cs.LG

Released Date: October 30, 2024

Authors: Chiu-Wai Yan¹, Shi Quan Foo¹, Van Hoan Trinh¹, Dit-Yan Yeung¹, Ka-Hing Wong², Wai-Kin Wong²

Aff.: ¹The Hong Kong University of Science and Technology; ²Hong Kong Observatory

Type	Model	Loss	Pixel-wise/Structural		Perceptual		Skill	Proposed
Type	Model	Loss	MAE $\downarrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FVD $\downarrow$	FSS $\uparrow$	RHD $\downarrow$
Pred.	ConvLSTM	MSE	196.4	0.6975	0.2538	451.5	0.6148	1.1504
	ConvLSTM	FACL	180.1	0.7463	0.1092	82.3	0.8172	0.3391
	PredRNN	MSE	173.8	0.7566	0.1875	337.8	0.7443	0.9559
	PredRNN	FACL	162.1	0.7812	0.0869	63.3	0.8549	0.3000
	SimVP	MSE	175.5	0.7547	0.1943	350.6	0.7275	0.9819
	SimVP	FACL	180.2	0.7394	0.1335	211.9	0.8168	0.3579
	Earthformer	MSE	171.5	0.7641	0.1828	320.3	0.7532	0.9407
	Earthformer	FACL	167.6	0.7768	0.0890	64.6	0.8463	0.3230
Gen.	LDCast^∗	-	234.0	0.7053	0.1541	110.7	0.6645	0.4343
	MCVD	-	219.8	0.7125	0.1033	44.7	0.7184	0.3941
	STRPM	-	154.0	0.7912	0.1017	117.4	0.8337	0.3216
	PreDiff	-	190.2	0.7570	0.0709	30.8	0.7975	0.3052
* The experiment setting for LDCast is changed to 8-in-8-out due to model constraints.