1 Introduction
1.1 Contribution
2 Related work
2.1 SfM and VisualSLAM
2.2 Feature points
2.3 CNN features
3 The pipeline: SfM using dense CNN features with keypoint relocalization
3.1 Dense feature extraction
3.2 Tentative matching
3.3 Keypoint relocalization
3.4 Feature verification using RANSAC with multiple homographies
3.5 3D reconstruction
4 Experiments
4.1 Results on Aachen Day-Night dataset
DoG+ | DenseCNN | DenseCNN | |
---|---|---|---|
RootSIFT [6] | w/o reloc | w/ reloc (Ours) | |
Night | 48 | 95 |
96
|
Day | 1910 | 1924 |
1944
|
DoG+ | DenseCNN | DenseCNN | |
---|---|---|---|
RootSIFT [6] | w/o reloc | w/ reloc (Ours) | |
0.5m | 15.31 | 5.10 |
18.37
|
1.0m | 25.61 | 14.29 |
33.67
|
5.0m | 36.73 | 45.92 |
69.39
|
10.0m | 35.71 | 61.22 |
81.63
|
20.0m | 39.80 | 69.39 |
82.65
|