notesum.ai

Published at December 10

EDGE: Unknown-aware Multi-label Learning by Energy Distribution Gap Expansion

cs.CV

Released Date: December 10, 2024

Authors: Yuchen Sun¹, Qianqian Xu, Zitai Wang, Zhiyong Yang, Junwei He

Aff.: ¹Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, CAS, Beijing, China

Arxiv: http://arxiv.org/pdf/2412.07499v1

Refer to caption

Training	Inference	LSUN			Texture			iSUN			mAP $\downarrow$
Training	Inference	FPR95 $\downarrow$	AUROC $\uparrow$	AUPR $\uparrow$	FPR95 $\downarrow$	AUROC $\uparrow$	AUPR $\uparrow$	FPR95 $\downarrow$	AUROC $\uparrow$	AUPR $\uparrow$	mAP $\downarrow$
BCE	MaxLogit	1.08	99.63	99.90	11.19	96.97	99.44	29.80	89.93	96.02	73.79
	MSP	22.41	95.13	98.65	45.51	87.64	97.62	60.58	79.99	92.89
	ODIN	1.31	99.23	99.81	11.51	96.31	99.28	33.42	88.70	95.71
	Mahalanobis	10.18	98.02	99.54	18.21	96.28	99.42	97.78	63.25	88.26
	Energy Score	0.75	99.72	99.92	9.11	97.33	99.49	27.51	90.28	96.16
	ReAct	3.38	98.98	99.69	8.58	97.78	99.62	27.39	89.88	95.60
	ViM	0.27	99.91	99.98	5.76	98.31	99.71	36.41	89.01	96.04
	ASH	1.38	98.88	98.90	8.24	97.82	99.61	30.45	90.21	96.81
	MaxEnergy	1.08	99.63	99.90	11.19	96.97	99.44	29.80	89.83	96.02
	JointEnergy	0.48	99.80	99.95	7.20	98.15	99.71	26.05	92.04	97.47
OE	MaxEnergy	5.79	98.78	99.72	9.38	98.09	99.72	25.42	93.41	97.95	68.47
OE	JointEnergy	69.19	65.80	81.33	60.30	71.41	89.58	57.67	78.84	88.62	68.47
EnergyOE	MaxEnergy	18.01	97.28	99.34	27.85	92.92	98.69	49.29	86.36	95.99	72.37
EnergyOE	JointEnergy	33.51	81.38	88.01	38.60	80.49	92.37	46.29	81.52	89.35	72.37
EDGE	MaxEnergy	0.07	99.95	99.99	4.18	98.90	99.82	20.15	93.90	97.84	73.50
EDGE	JointEnergy	0.01	99.91	99.98	3.09	99.17	99.88	18.45	94.86	98.40	73.50