SENetCount: An Optimized Encoder-Decoder Architecture with Squeeze-and-Excitation for Crowd Counting

Table 2

ResNet, SE-ResNet, and SE-ResNeXt backbone network.

Operate	Output size	ResNet50	SE-ResNet50	SE-ResNeXt50	SE-ResNeXt101
Image	1	—
Conv2d	1/2	, 64, stride 2
Max pooling	1/4	, stride 2
Bottleneck1	1/4
Bottleneck2	1/8
Bottleneck3	1/16
Bottleneck4	1/32

ResNet₅₀ (Left). SE-ResNet₅₀ (Middle). SE-ResNeXt_50/101 (Right). The patterns and operations with definite residual building block parameter settings are recorded inside the brackets. The statistic of parallel or cascade blocks in a stage is given outside the brackets. The interior bracket next to denotes the output dimension of the two fully-connected layers in a SE block. suggests grouped-connected layers with 32 groups.