Submission details
| Method name | DenseVLAD (top-20) & D2-Net - camera triplets |
| Dataset | robotcar |
| Submission date | June 17, 2020, 12:42 p.m. |
| Last modified | June 17, 2020, 12:46 p.m. |
| Publication url | http://openaccess.thecvf.com/content_CVPR_2019/papers/Dusmanu_D2-Net_A_Trainable_CNN_for_Joint_Description_and_Detection_of_CVPR_2019_paper.pdf |
| Code url | https://github.com/mihaidusmanu/d2-net |
| Other info | Uses DenseVLAD (single-scale) to identify the top-20 closest database images, then uses single-scale D2-Net features for feature matching. Instead of estimating the camera poses of individual images, camera triplets consisting of images taken at the same point in time are used. |
Detailed results for all conditions
| Method | night all | night rain | overcast winter | sun | rain | snow | dawn | dusk | night | day all | overcast summer |
|---|---|---|---|---|---|---|---|---|---|---|---|
| DenseVLAD (top-20) & D2-Net - camera triplets | 28.6 / 50.0 / 67.5 | 34.3 / 55.2 / 71.8 | 48.5 / 80.5 / 96.2 | 48.5 / 77.2 / 94.3 | 60.1 / 83.1 / 96.9 | 58.5 / 85.9 / 95.7 | 55.1 / 81.4 / 95.7 | 62.9 / 83.2 / 95.9 | 22.8 / 44.7 / 63.2 | 53.0 / 81.5 / 96.3 | 38.7 / 79.3 / 99.1 |