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

<div>(a) The stacked atrous convolutional feature layers with the sizeable atrous rate of 6 and 12 lose a large amount of pixel information. (b) The stacked atrous convolutional feature layers with the dense atrous rate of 1, 2, or 3 in the proposed DASPP cover all pixel information. The orange color indicates the source of information.</div>

Wireless Communications and Mobile Computing

fig4

Figure 4

Figure 4: SENetCount: An Optimized Encoder-Decoder Architecture with Squeeze-and-Excitation for Crowd Counting 