Viewpoint robust knowledge distillation for accelerating vehicle re-identification

EURASIP Journal on Advances in Signal Processing

Table 3 The role (%) of viewpoint robust knowledge distillation on the VeRi776 [1] dataset and the VehicleID [2] dataset

Student networks	Using VRKD	Training tricks		VeRi776		VehicleID
						Test800		Test1600		Test2400
		LS=1	LSR	MAP	R1	MAP	R1	MAP	R1	MAP	R1
CNN	No	No	Yes	42.16	78.78	67.67	61.35	65.83	60.09	62.66	56.68
	Yes	No	Yes	52.20	83.02	77.17	69.69	73.98	67.51	70.59	63.86
SDC-CNN [14]	No	No	Yes	58.50	89.81	78.28	71.22	75.34	69.23	72.41	66.37
	Yes	No	Yes	69.87	93.62	83.44	76.37	79.62	72.87	76.39	69.56
DenseNet121 [16]	No	No	Yes	54.26	85.64	78.84	72.22	75.84	69.70	73.01	66.98
	Yes	No	Yes	65.22	90.11	84.67	78.00	80.67	74.19	78.03	71.94
ResNet-18 [15]	No	No	Yes	68.03	92.37	84.04	77.51	79.93	73.39	77.40	70.74
	Yes	No	Yes	74.67	94.52	87.29	81.29	83.35	77.21	81.30	74.62
ResNet-50 [15]	No	No	Yes	73.98	95.05	87.36	81.31	83.50	77.27	80.94	74.35
	Yes	No	Yes	75.74	94.58	88.70	82.26	84.14	77.90	81.59	75.18