EURASIP Journal on Image and Video Processing

Table 3 Performance of pruning techniques from the literature in terms of rank-1 accuracy and computational complexity (memory(M): number of parameters and time (T): time required for one forward pass). To ease comparison, we include the out results produced with ThiNet (channel pruning method)

From: Exploiting prunability for person re-identification

Dataset	ResNet56 trained on CIFAR10
Algorithm	Original			Pruned
	rank-1	T	M	rank-1	T	M
L1 [33]	93.04	0.125	0.85	93.06	0.091	0.73
Auto-Balanced [70]	93.93	0.142	N/D	92.94	0.055	N/D
Redundant channel [67]	93.39	0.125	0.85	93.12	0.091	0.65
Play and Prune [69]	93.39	0.125	0.85	93.09	0.039	N/D
FPGM [68]	93.39	0.125	0.85	92.73	0.059	N/D
Dataset	VGG16 trained on ImageNet
Algorithm	Original			Pruned
	rank-1	T	M	rank-1	T	M
ThiNet [71]	90.01	30.94	138.34	89.41	9.58	131.44
Taylor [32]	89.30	30.96	N/D	87.06	11.5	N/D
HaoLi [33]	90.01	30.94	138.34	89.13	9.58	130.87
Channel Pruning [35]	90.01	30.94	138.34	88.10	7.03	131.44
Dataset	ResNet50 trained on ImageNet
Algorithm	Original			Pruned
	rank-1	T	M	rank-1	T	M
Entropy [34]	72.88	3.86	25.56	70.84	2.52	17.38
ThiNet [71]	75.30	7.72	25.56	72.03	3.41	138.00
FPGM [68]	75.30	7.72	25.56	74.83	3.58	N/D

Back to article page