notesum.ai

Published at December 5

ONER: Online Experience Replay for Incremental Anomaly Detection

cs.CV

Released Date: December 5, 2024

Authors: Yizhou Jin¹, Jiahui Zhu, Guodong Wang, Shiwei Li, Jinjin Zhang, Qingjie Liu, Xinyue Liu, Yunhong Wang

Aff.: ¹Beihang University, Beijing, China

Arxiv: http://arxiv.org/pdf/2412.03907v1

Refer to caption

DRAEM
Method	Pixel-level				Image-level
Method	ACC $\uparrow$		FM $\downarrow$		ACC $\uparrow$		FM $\downarrow$
	AUROC	AUPR	AUROC	AUPR	AUROC	AUPR	AUROC	AUPR
Unified Finetuning	61.58	11.73	34.27	53.37	57.74	78.73	39.45	20.05
$\text{Finetuning with Sample Replay}^{\#}$	85.29	36.29	10.56	29.01	76.01	87.89	21.18	10.89
$\text{Incremental Classifier-based}^{\#*}$	82.67	29.71	13.18	35.59	80.62	87.11	16.57	11.67
BGAD
Unified Finetuning	50.36	6.61	46.07	48.01	51.76	73.85	42.10	22.64
$\text{Finetuning with Sample Replay}^{\#}$	91.29	32.95	5.14	21.67	79.85	89.46	14.01	7.03
$\text{Incremental Classifier-based}^{\#*}$	84.77	24.53	11.66	30.09	83.39	89.32	10.47	7.17
PatchCore*	89.72	22.15	3.83	19.33	90.35	94.94	-0.28	0.31
UCAD*	90.11	21.42	6.10	24.16	90.74	95.29	1.80	0.85
IUF	79.01	15.95	5.65	6.22	75.08	86.37	1.17	1.61
CAD	-	-	-	-	87.06	93.32	1.89	1.66
CAD + PANDA	-	-	-	-	47.83	74.07	36.04	18.47
CAD + CutPaste	-	-	-	-	74.61	86.31	17.18	10.00
ONER (Ours)	94.53	49.66	0.91	1.33	93.36	97.26	0.59	0.15