SENetCount: An Optimized Encoder-Decoder Architecture with Squeeze-and-Excitation for Crowd Counting
Table 4
Fundament experiments.
Networks
Part_A
Part_B
MAE
RMSE
MAE
RMSE
ResNetCount50,3W/O
121.4
178.0
11.0
17.6
ResNetCount50,3
78.3
125.8
8.2
13.6
ResNetCount50,4
78.0
129.7
8.1
13.7
SE-ResNetCount50,3W/O
111.9
177.6
11.1
17.3
SE-ResNetCount50,3
76.0
120.5
7.7
12.2
SE-ResNetCount50,4
76.8
123.6
8.2
13.5
SE-ResNeXtCount50,3W/O
114.9
176.4
11.2
15.9
SE-ResNeXtCount50,3
71.9
118.7
8.0
13.3
SE-ResNeXtCount50,4
72.1
121.8
8.1
13.2
3 indicates that only the first three bottlenecks are selected, and 4 indicates that all the four bottlenecks are chosen. W/O means that the pretraining strategy is not used.
