An Enhanced Visual Attention Siamese Network That Updates Template Features Online

Table 1

Network structure and operations corresponding to each network block (block represents network block, Gold-SPool represents golden stochastic pooling, dilation represents dilated convolution, ResNet in Figure 1 includes Block1, Block2, Block3, Block4, and Block5, and “—” represents no operation).

Block	Operation	Template size	Search size
—	—	127 × 127 × 3	255 × 255 × 3
Block1	7 × 7, 64, 3 × 3Gold-SPool, s = 2	31 × 31 × 64	62 × 62 × 64
Block2		15 × 15 × 256	31 × 31 × 256
Block3 + dilation		15 × 15 × 512	31 × 31 × 512
Attention	—	15 × 15 × 256	31 × 31 × 256
Block4 + dilation		15 × 15 × 1024	31 × 31 × 1024
Attention	—	15 × 15 × 512	31 × 31 × 512
Block5 + dilation		15 × 15 × 2048	31 × 31 × 2048
Attention	—	15 × 15 × 1024	31 × 31 × 1024