Viewpoint robust knowledge distillation for accelerating vehicle re-identification

Xie, Yi; Shen, Fei; Zhu, Jianqing; Zeng, Huanqiang

doi:10.1186/s13634-021-00767-x

EURASIP Journal on Advances in Signal Processing

Table 1 The performance (%) comparison with the state-of-the-art methods on the VeRi776 [1] dataset and the VehicleID [2] dataset. - denotes the result is not provided by the reference

From: Viewpoint robust knowledge distillation for accelerating vehicle re-identification

Methods	Testing Models	Parameters	Computations	Resolutions	Training Tricks			VeRi776		VehicleID
		(Millions)	(G FLOPs)
										Test800		Test1600		Test2400
					LS=1	LSR	TRI	MAP	R1	MAP	R1	MAP	R1	MAP	R1
XGAN [7]	3 CNN	-	-	128×128	No	No	No	24.65	60.20	-	52.89	-	-	-	-
SCCN [4]	9 CNN+Bi-LSTM	-	-	100×100	No	No	No	25.12	60.83	-	48.63	-	-	-	-
ABLN [3]	1 CNN+Bi-LSTM	-	-	128×128	No	No	No	-	58.14	-	52.63	-	-	-	-
VAMI [5]	1 CNN+MLP	-	-	256×256	No	No	No	50.13	77.03	-	63.12	-	52.87	-	47.34
VRKD	1 CNN	1.01	0.41	128×128	No	No	No	52.16	83.19	77.64	70.24	74.49	67.83	71.20	64.41
VRKD	1 CNN	1.01	0.41	128×128	No	Yes	No	52.20	83.02	77.17	69.69	73.98	67.51	70.59	63.86
JQD³Ns [9]	4 SDC-CNNs	44.41	36.81	128×128	No	No	No	61.30	89.69	-	-	-	-	-	-
QD-DLF [8]	4 SDC-CNNs	44.41	36.81	128×128	No	No	No	61.83	88.50	76.54	72.32	74.63	70.66	68.41	64.14
VRKD	1 SDC-CNN	10.86	9.2	128×128	No	No	No	69.59	93.44	83.77	76.89	80.08	73.40	76.64	69.75
VRKD	1 SDC-CNN	10.86	9.2	128×128	No	Yes	No	69.87	93.62	83.44	76.37	79.62	72.87	76.39	69.56
AAVER [10]	ResNet-101	47.88	13.03	224×224	-	No	No	61.18	88.97	-	74.69	-	68.62	63.54	-
PAMTRI [11]	DenseNet121	8.8	3.76	256×256	-	No	Yes	71.88	92.86	-	-	-	-	-	-
EALN [12]	ResNet-50	28.89	6.25	224×224	-	No	Yes	-	-	77.5	75.11	74.2	71.78	71.0	69.30
PartReg [13]	ResNet-50	28.89	8.16	256×256	-	No	No	70.2	92.2	-	78.4	-	75.0	-	74.2
PartReg [13]	ResNet-50	28.89	32.63	512×512	-	No	No	74.3	94.3	-	-	-	-	-	-
VehicleNet [17]	ResNet-50	28.89	16.32	256×256	-	No	No	83.41	96.78	-	83.64	-	81.35	-	79.46
VRKD	ResNet-18	11.46	3.99	256×256	No	No	No	74.40	94.52	87.28	81.37	83.32	77.03	80.87	74.46
VRKD	ResNet-18	11.46	15.97	512×512	No	No	No	76.29	94.64	89.22	83.86	85.43	79.65	83.26	77.23
VRKD	ResNet-18	11.46	3.99	256×256	No	Yes	No	74.48	94.64	87.26	81.21	83.14	76.80	80.65	74.13
VRKD	ResNet-18	11.46	3.99	256×256	Yes	Yes	No	76.12	95.65	87.68	81.77	83.52	77.23	81.13	74.73
VRKD	ResNet-18	11.46	15.97	512×512	Yes	Yes	No	76.39	94.58	89.22	83.86	85.46	79.72	83.35	77.34

⁺These state-of-the-art methods’ results come from their original literature.

Back to article page