📊 Results

🌍 Open World Learning

Table 1. The experiments of OWL on OpenEarthSensing dataset. ID Acc and ODD Acc are the in-distribution and out-of-distribution performance, respectively, and Avg denotes the average performance of each session.

OOD Method	CIL Method	ID Acc	OOD Acc	Session 1	Session 2	Avg
MSP	LwF	91.17	55.01	91.27	42.11	66.69
	EWC	91.17	55.01	91.27	28.89	60.08
	iCaRL	91.17	55.01	91.27	50.29	70.78
MLS	LwF	90.6	63.85	91.27	44.09	67.68
	EWC	90.6	63.85	91.27	29.83	60.55
	iCaRL	90.6	63.85	91.27	51.33	71.30
VIM	LwF	93.5	59.99	91.27	43.49	67.38
	EWC	93.5	59.99	91.27	31.33	61.30
	iCaRL	93.5	59.99	91.27	33.78	62.52

🌱 Incremental Learning

Fig 1. Experimental results of CIL. The left figure presents results in randomized order, while the right figure displays systematically organized results arranged from coarse to fine granularity.

CIL

🧩 Few-Shot Class-Incremental Learning

Table 2. The experimental results of few-shot class-incremental learning on the OpenEarthSensing dataset. Shots denote the training samples for each category.

	50-shot		10-shot		5-shot		1-shot
	Last	Avg	Last	Avg	Last	Avg	Last	Avg
Alice	59.54	64.66	59.17	68.64	58.82	68.35	58.94	68.4
FACT	46.42	49.21	46.38	49.15	46.36	49.15	46.37	49.15
SAVC	71.71	79.55	72.23	80.07	66.61	75.17	59.63	76.77

🚨 Out-of-Distribution Detection

CNN-based Methods

Table 3. OOD detection performance on OES benchmark. 'Near' represents the average AUROC for Near-OOD datasets, 'Far' indicates the average AUROC for Far-OOD datasets.

Method	Standard		Res Bias		Aerial		MSRGB		IR
	Near	Far	Near	Far	Near	Far	Near	Far	Near	Far
MSP	88.42	93.91	66.51	78.4	54.38	56.85	65.50	66.92	61.47	65.35
ODIN	87.14	95.79	67.09	75.2	52.85	57.04	66.55	61.55	62.11	73.28
MDS	83.15	96.54	53.86	84.71	48.79	54.76	66.31	81.78	83.64	57.74
MLS	88.59	96.12	66.44	83.17	53.93	59.78	64.46	63.37	62.49	67.06
VIM	90.35	98.75	60.33	83.93	50.69	59.72	64.90	81.75	57.65	51.08
FBDB	90.24	98.17	66.64	87.87	54.49	60.41	59.45	74.49	61.40	68.62
VOS	86.19	95.68	63.37	81.32	51.10	60.01	59.77	58.72	59.47	60.26
LogiNorm	89.00	95.15	68.80	80.29	53.25	56.72	77.69	55.43	64.12	63.97
DML	84.38	90.36	65.78	76.16	52.89	58.60	62.56	50.68	60.39	50.56

CLIP-based Methods

Table 4. CLIP based methods' performance on OES benchmark. 'Near' represents the average AUROC for Near-OOD datasets, 'Far' indicates the average AUROC for Far-OOD datasets.

Method	Standard		Res Bias		Aerial		MSRGB		IR
	Near	Far	Near	Far	Near	Far	Near	Far	Near	Far
MaxLogits	53.31	43.95	68.99	63.32	64.73	40.46	68.22	9.34	62.73	37.00
MCM	61.79	52.60	59.97	51.94	66.07	67.85	58.90	55.89	54.41	40.43
GLMCM	62.07	52.33	59.20	51.57	65.20	67.42	57.32	56.89	51.75	42.30
CoOp	86.04	94.21	64.09	73.36	61.22	76.40	66.73	90.22	61.30	45.16
LoCoOp	85.71	90.94	66.20	71.67	64.18	76.52	69.64	86.28	61.41	43.33
SCT	85.56	90.78	65.37	70.30	64.14	77.67	68.58	86.41	60.81	42.48
DPM	91.19	99.24	68.60	92.61	60.50	71.26	74.66	93.56	65.11	75.10