notesum.ai

Published at December 6

SAMCL: Empowering SAM to Continually Learn from Dynamic Domains

cs.CV

Released Date: December 6, 2024

Authors: Zeqing Wang¹, Kangye Ji¹, Di Wang¹, Fei Cheng¹

Aff.: ¹Xidian University

Arxiv: http://arxiv.org/pdf/2412.05012v1

Refer to caption

Kvasir $\rightarrow$ CAMO $\rightarrow$ ISTD $\rightarrow$ ISIC $\rightarrow$ COD
Method	AA			FM			FT
Method	mIoU $\uparrow$	mF1 $\uparrow$	mMAE $\downarrow$	mIoU $\downarrow$	mF1 $\downarrow$	mMAE $\uparrow$	mIoU $\uparrow$	mF1 $\uparrow$	mMAE $\downarrow$
SAM-Adapter [7]	0.428	0.565	0.125	0.191	0.190	-0.046	0.318	0.457	0.196
SAM-LST [5]	0.421	0.533	0.170	0.239	0.222	-0.180	0.291	0.421	0.221
AutoSAM [30]	0.184	0.273	0.258	0.330	0.330	-0.103	0.124	0.199	0.423
Lora [14]	0.703	0.804	0.064	0.123	0.089	-0.034	0.513	0.521	0.158
EWC [17]	0.716	0.816	0.058	0.111	0.078	-0.028	0.549	0.663	0.160
ER [6]	0.808	0.881	0.035	0.010	0.007	-0.003	0.630	0.748	0.087
DER [4]	0.804	0.879	0.035	0.022	0.015	-0.005	0.643	0.760	0.082
SPPA [23]	0.282	0.407	0.149	0.337	0.315	-0.072	0.417	0.550	0.197
LAG [40]	0.703	0.810	0.063	0.099	0.066	-0.025	0.452	0.576	0.205
O-LoRA [34]	0.704	0.806	0.059	0.091	0.066	-0.023	0.519	0.642	0.160
SAMCL (Ours)	0.832	0.898	0.029	0.005	0.004	-0.001	0.668	0.775	0.084