Feature Point Detection and Description Networks Based on Asymmetric Convolution and the Cross-ResolutionImage-Matching Method
Table 5
HPatches ablation experiment.
Homography estimation
Detector metrics
Descriptor metrics
= 1↑
= 3↑
= 5↑
Rep.↑
MLE.↓
NNmAP↑
M.score↑
Conv + RELU + max pooling
0.486
0.764
0.821
0.644
1.093
0.800
0.490
ACB + RELU + max pooling
0.495
0.775
0.830
0.647
1.074
0.806
0.491
Conv + ELU + max pooling
0.497
0.771
0.824
0.645
1.091
0.811
0.490
Conv + RELU + mix pooling
0.500
0.765
0.827
0.648
1.072
0.809
0.500
ACB + ELU + mix pooling
0.548
0.778
0.845
0.656
1.024
0.845
0.534
The meanings in bold in Table 5 represent the values of the best performing methods for the indicators in this column. Meanwhile, the method proposed in this paper has the best performance in all indexes in this table.
