A Distinguishable Pseudo-Feature Synthesis Method for Generalized Zero-Shot Learning

Table 2

Quantitative comparisons of average per-class GZSL classification accuracy (%).

Method		AWA2			aPY			CUB			SUN
		†	LATEM [9]	77.3	11.5	20.0	73.0	0.1	0.2	57.3	15.2	24.0	28.8	14.7	19.5
DEM [19]	86.4		30.5	45.1	75.1	11.1	19.4	54.0	19.6	13.6	34.3	20.5	25.6
CPL [8]	83.1		51.0	63.2	73.2	19.6	30.9	58.6	28.0	37.9	32.4	21.9	26.1
DVBE [20]	70.8		63.6	67.0	58.3	32.6	41.8	60.2	53.2	56.5	37.2	45.0	40.7
DCC [21]	82.9		55.1	66.2	74.8	34.4	47.2	57.7	46.5	51.5	41.0	33.1	36.6
HSVA [22]	79.3		57.8	66.9	—	—	—	59.5	51.9	55.5	39.0	48.6	43.3
†	SEZEL [23]	68.1	58.3	62.8	—	—	—	53.3	41.5	46.7	30.5	40.9	34.9
	DUET [10]	90.2	48.2	63.4	55.6	21.8	31.3	80.1	39.7	53.1	—	—	—
	Inf-FG [26]	63.4	58.3	60.7	—	—	—	57.0	45.8	50.8	37.1	44.7	40.5
	LDMS [25]	71.8	60.9	65.9	66.3	37.4	47.8	61.6	48.0	53.9	36.2	45.6	40.3
	FREE [27]	75.4	60.4	67.1				59.9	55.7	57.7	37.7	44.8	40.9
	GCF [28]	75.1	60.4	67.0	56.8	37.1	44.9	59.7	61.0	60.3	37.8	47.9	42.2
†	SPF [15]	60.9	52.4	56.3	—	—	—	63.4	30.2	40.9	59.0	32.2	41.6
	TCN [14]	65.8	61.2	63.4	64.0	24.1	35.1	52.0	52.6	52.3	37.3	31.2	34.0
	LIUF [16]	83.5	60.6	70.2	79.1	38.2	51.6	54.0	51.2	52.5	40.4	45.7	42.9
	DPFS	87.3	61	71.8	83.0	38.6	52.7	63.8	54.0	58.5	43.0	49.6	46.1