Image Geolocation Method Based on Attention Mechanism Front Loading and Feature Fusion
Table 3
The performance comparisons between the proposed method and NetVLAD on the Pitts 30k [81].
Noise filtering layer
Feature aggregation layer
PCA
Dimension
R@1
R@5
R@10
R@20
×
NetVLAD
×
32768
79.45
90.10
92.77
95.19
×
NetVLAD
√
512
77.52
89.41
92.59
95.39
×
NetVLAD
√
1024
78.92
90.10
92.87
95.48
×
NetVLAD
√
2048
79.55
90.42
92.97
95.38
×
NetVLAD
√
4096
79.50
90.23
92.9
95.19
√
NetVLAD
×
32768
80.99
91.09
93.57
95.48
√
NetVLAD
√
512
80.34
91.37
93.97
95.76
√
NetVLAD
√
1024
81.24
91.61
94.07
95.77
√
NetVLAD
√
2048
81.41
91.62
93.85
95.85
√
NetVLAD
√
4096
81.29
91.37
93.63
95.6
√
NetVLAD + SPP+GeM
×
48128
83.67
92.36
94.16
95.77
√
NetVLAD + SPP+GeM
√
512
82.88
92.39
94.91
96.24
√
NetVLAD + SPP+GeM
√
1024
83.73
92.81
94.91
96.20
√
NetVLAD + SPP+GeM
√
2048
84.01
92.78
94.81
96.11
√
NetVLAD + SPP+GeM
√
4096
83.92
92.62
94.5
95.95
“×” means that the operation corresponding to the column name is not applied to the model, and “√” means the opposite. “Dimension” denotes the dimension of the final descriptor; “R@N” denotes .
