Benchmark
Listed below are the public results on the three benchmark datasets. The localization results are reported as the percentage of query images which where localized within three given translation and rotation thresholds, for each condition.
Aachen Day-Night dataset
Localization thresholds:
- Day: (0.25m, 2°) / (0.5m, 5°) / (5m, 10°)
- Night: (0.5m, 2°) / (1m, 5°) / (5m, 10°)
| Method | day | night | Last updated |
|---|---|---|---|
| Visual Localization Using Sparse Semantic 3D Map | 71.8 / 91.5 / 96.8 | 40.8 / 63.3 / 80.6 | June 2, 2019, 4 a.m. |
| Hierarchical-Localization NetVLAD+SuperPoint | 80.5 / 87.4 / 94.2 | 42.9 / 62.2 / 76.5 | June 1, 2019, 9:50 a.m. |
| Hierarchical-Localization (multi-camera when available) | 80.5 / 87.4 / 94.2 | 42.9 / 62.2 / 76.5 | June 2, 2019, 6:55 a.m. |
| DenseVLAD & D2-Net (top-20) | 80.1 / 88.0 / 93.4 | 39.8 / 55.1 / 74.5 | May 29, 2019, 7:48 a.m. |
| UR2KID Scape Technologies (paper coming soon, not scene specific, single scale) | 0.0 / 0.0 / 0.0 | 46.9 / 67.3 / 88.8 | July 3, 2019, 3:39 p.m. |
| D2-Net - single-scale | 0.0 / 0.0 / 0.0 | 45.9 / 68.4 / 88.8 | July 16, 2019, 12:21 p.m. |
| R2D2 10k keypoints | 0.0 / 0.0 / 0.0 | 45.9 / 66.3 / 88.8 | June 17, 2019, 9:56 a.m. |
| Asymmetric Hypercolumn Matching | 47.8 / 72.2 / 91.3 | 30.6 / 53.1 / 78.6 | June 5, 2019, 7:28 p.m. |
| densecontextdesc10k_upright_mixedmatcher | 0.0 / 0.0 / 0.0 | 46.9 / 65.3 / 87.8 | July 16, 2019, 11:13 a.m. |
| D2-Net - single-scale off-the-shelf | 0.0 / 0.0 / 0.0 | 41.8 / 69.4 / 86.7 | July 16, 2019, 4:57 p.m. |
| DELF - new model | 0.0 / 0.0 / 0.0 | 39.8 / 61.2 / 85.7 | June 12, 2019, 3:52 p.m. |
| NetVLAD+SuperPoint top 10 (baseline) | 33.9 / 48.8 / 78.5 | 25.5 / 45.9 / 80.6 | June 1, 2019, 8:59 p.m. |
| CityScaleLocalization | 52.3 / 80.0 / 94.3 | 24.5 / 33.7 / 49.0 | May 10, 2019, 2:22 p.m. |
| DELF - old model | 0.0 / 0.0 / 0.0 | 39.8 / 60.2 / 84.7 | May 21, 2019, 2:52 p.m. |
| Localizing Visual Landmarks for 2D matching | 62.4 / 71.8 / 79.9 | 24.5 / 35.7 / 44.9 | April 20, 2019, 5:21 p.m. |
| saliency_ranking_net (multi-scale) | 0.0 / 0.0 / 0.0 | 44.9 / 59.2 / 77.6 | April 29, 2019, 2:33 p.m. |
| Active Search | 57.3 / 83.7 / 96.6 | 19.4 / 30.6 / 43.9 | March 13, 2019, 9 a.m. |
| contextdesc10k_upright | 0.0 / 0.0 / 0.0 | 37.8 / 56.1 / 80.6 | July 16, 2019, 11:17 a.m. |
| DGCNCCC | 22.9 / 49.8 / 84.7 | 19.4 / 37.8 / 68.4 | May 30, 2019, 10:17 a.m. |
| HesAffNet-HardNet2 | 0.0 / 0.0 / 0.0 | 37.8 / 54.1 / 75.5 | May 31, 2019, 10:35 p.m. |
| Upright RootSIFT (Feature Challenge Baseline) | 0.0 / 0.0 / 0.0 | 33.7 / 52.0 / 65.3 | May 2, 2019, 8:57 a.m. |
| ELF | 0.0 / 0.0 / 0.0 | 13.3 / 21.4 / 30.6 | May 3, 2019, 8:36 p.m. |
| DenseVLAD | 0.0 / 0.1 / 22.8 | 0.0 / 2.0 / 14.3 | April 11, 2019, 5:43 p.m. |
| NetVLAD | 0.0 / 0.2 / 18.9 | 0.0 / 2.0 / 12.2 | May 31, 2019, 11:16 p.m. |
| Localizing Visual Landmarks for Place Recognition | 0.0 / 0.2 / 20.8 | 0.0 / 1.0 / 10.2 | June 2, 2019, 10:06 p.m. |
| DGCNCCC | 0.0 / 0.1 / 9.5 | 0.0 / 0.0 / 8.2 | March 17, 2019, 11:42 a.m. |
| NetVlad_NN_Pose | 0.0 / 0.1 / 7.0 | 0.0 / 0.0 / 4.1 | March 17, 2019, 11:46 a.m. |
| joint image segemation and depth to posenet | 0.0 / 0.1 / 24.9 | 0.0 / 0.0 / 0.0 | May 31, 2019, 11:45 a.m. |
| FAB-MAP | 0.0 / 0.0 / 4.6 | 0.0 / 0.0 / 0.0 | March 13, 2019, 11:54 a.m. |
CMU Seasons dataset
Localization thresholds:
- All conditions: (0.25m, 2°) / (0.5m, 5°) / (5m, 10°)
| Method | urban | suburban | park | Last updated |
|---|---|---|---|---|
| Hierarchical-Localization NetVLAD+SuperPoint | 91.7 / 94.6 / 97.7 | 74.5 / 81.5 / 91.3 | 54.3 / 62.5 / 79.0 | June 1, 2019, 9:50 a.m. |
| LocalSfM (Upright RootSIFT) (has partial access to ground truth) | 72.8 / 74.1 / 76.1 | 55.2 / 57.7 / 61.3 | 41.8 / 44.5 / 48.7 | March 13, 2019, 11:42 a.m. |
| Semantic Match Consistency | 75.0 / 82.1 / 87.8 | 44.0 / 53.6 / 63.7 | 30.0 / 37.9 / 48.2 | March 18, 2019, 11:47 p.m. |
| EffecientPairWise2D-3DMatching | 64.1 / 67.8 / 71.5 | 30.2 / 35.5 / 40.9 | 23.8 / 28.1 / 34.0 | March 20, 2019, 9:52 a.m. |
| DenseVLAD | 22.2 / 48.6 / 92.8 | 9.8 / 26.6 / 85.2 | 10.3 / 27.1 / 77.0 | April 15, 2019, 11:03 a.m. |
| Active Search | 55.2 / 60.3 / 65.1 | 20.7 / 25.9 / 29.9 | 12.7 / 16.3 / 20.8 | March 13, 2019, 9:04 a.m. |
| NetVLAD | 17.4 / 40.3 / 93.2 | 7.6 / 21.0 / 80.5 | 5.6 / 15.7 / 65.8 | April 11, 2019, 9:16 p.m. |
| CityScaleLocalization | 36.7 / 42.0 / 53.1 | 8.6 / 11.7 / 21.1 | 7.0 / 9.6 / 17.0 | March 13, 2019, 11:59 a.m. |
| FAB-MAP | 2.7 / 6.4 / 27.3 | 0.5 / 1.5 / 13.6 | 0.8 / 1.7 / 11.5 | March 13, 2019, 11:49 a.m. |
Extended CMU Seasons dataset
Localization thresholds:
- All conditions: (0.25m, 2°) / (0.5m, 5°) / (5m, 10°)
| Method | urban | suburban | park | Last updated |
|---|---|---|---|---|
| Hierarchical-Localization (multi-camera when available) | 91.6 / 96.4 / 99.1 | 84.7 / 91.5 / 98.6 | 69.3 / 77.8 / 90.5 | June 2, 2019, 6:55 a.m. |
| Visual Localization Using Sparse Semantic 3D Map | 88.8 / 93.6 / 96.3 | 78.0 / 83.8 / 89.2 | 63.6 / 70.3 / 77.3 | June 2, 2019, 4 a.m. |
| Hierarchical-Localization NetVLAD+SuperPoint | 89.5 / 94.2 / 97.9 | 76.5 / 82.7 / 92.7 | 57.4 / 64.4 / 80.4 | June 1, 2019, 9:50 a.m. |
| Asymmetric Hypercolumn Matching | 65.7 / 82.7 / 91.0 | 66.5 / 82.6 / 92.9 | 54.3 / 71.6 / 84.1 | June 5, 2019, 7:24 p.m. |
| Localizing Visual Landmarks for 2D matching | 84.3 / 89.3 / 93.0 | 68.0 / 75.1 / 84.4 | 42.4 / 51.4 / 69.7 | April 25, 2019, 8:12 a.m. |
| CityScaleLocalization | 71.2 / 74.6 / 78.7 | 57.8 / 61.7 / 67.5 | 34.5 / 37.0 / 42.2 | March 8, 2019, 8:20 p.m. |
| Localizing Visual Landmarks for Place Recognition | 17.3 / 42.5 / 89.0 | 5.8 / 19.4 / 76.1 | 6.6 / 23.1 / 73.0 | June 2, 2019, 10:07 p.m. |
| DGCNCCC | 17.1 / 41.5 / 89.1 | 8.9 / 26.8 / 77.1 | 4.8 / 16.2 / 63.3 | June 2, 2019, 1:27 p.m. |
| Netvlad_localization | 16.4 / 38.5 / 78.9 | 8.8 / 27.1 / 77.0 | 4.7 / 16.0 / 63.2 | June 25, 2019, 9:44 p.m. |
| DenseVLAD | 14.7 / 36.3 / 83.9 | 5.3 / 18.7 / 73.9 | 5.2 / 19.1 / 62.0 | March 12, 2019, 3:21 p.m. |
| NetVLAD | 12.2 / 31.5 / 89.8 | 3.7 / 13.9 / 74.7 | 2.6 / 10.4 / 55.9 | June 2, 2019, 9:12 a.m. |
| joint image segemation and depth to posenet | 0.3 / 1.5 / 31.9 | 0.1 / 0.3 / 12.5 | 0.0 / 0.2 / 9.0 | May 31, 2019, 11:46 a.m. |
RobotCar Seasons dataset
Localization thresholds:
- All conditions: (0.25m, 2°) / (0.5m, 5°) / (5m, 10°)
| Method | day all | night all | Last updated |
|---|---|---|---|
| Visual Localization Using Sparse Semantic 3D Map | 54.5 / 81.6 / 96.7 | 12.3 / 28.5 / 46.5 | June 2, 2019, 4 a.m. |
| Hierarchical-Localization (multi-camera when available) | 53.8 / 80.4 / 96.0 | 11.2 / 27.7 / 49.1 | June 2, 2019, 6:54 a.m. |
| Semantic Match Consistency | 50.3 / 79.3 / 95.2 | 7.1 / 22.4 / 45.3 | March 18, 2019, 11:47 p.m. |
| Active Search on sequences of camera triplets (uses ground truth relative poses) | 46.6 / 80.1 / 97.0 | 5.8 / 21.0 / 43.1 | March 13, 2019, 12:03 p.m. |
| Hierarchical-Localization NetVLAD+SuperPoint | 53.1 / 79.1 / 95.5 | 7.2 / 17.4 / 34.4 | June 1, 2019, 9:50 a.m. |
| Asymmetric Hypercolumn Matching | 45.7 / 78.0 / 95.1 | 22.3 / 61.8 / 94.5 | June 5, 2019, 7:25 p.m. |
| Localizing Visual Landmarks for 2D matching | 51.1 / 77.7 / 92.3 | 13.8 / 30.3 / 57.7 | May 16, 2019, 2:09 a.m. |
| EffecientPairWise2D-3DMatching | 42.6 / 78.2 / 95.7 | 8.5 / 14.1 / 20.6 | March 20, 2019, 9:52 a.m. |
| Scalable and Accurate Image-based Localization | 48.0 / 78.0 / 94.2 | 3.4 / 9.5 / 17.0 | May 9, 2019, 9:02 a.m. |
| Active Seach on Camera Triplets | 45.5 / 77.0 / 94.7 | 2.7 / 6.9 / 12.1 | March 13, 2019, 9:31 a.m. |
| CityScaleLocalization | 45.3 / 73.5 / 90.1 | 0.6 / 2.6 / 7.2 | March 13, 2019, 11:59 a.m. |
| Active Search | 35.6 / 67.9 / 90.4 | 0.9 / 2.1 / 4.3 | March 13, 2019, 9:04 a.m. |
| DenseVLAD (single-scale, top-1 interpolation) | 9.8 / 32.8 / 90.7 | 1.8 / 6.6 / 25.6 | May 29, 2019, 3:27 p.m. |
| ToDayGAN + DenseVLAD | 7.6 / 31.2 / 91.2 | 2.2 / 10.8 / 50.5 | May 27, 2019, 8:05 p.m. |
| DenseVLAD (single-scale) | 8.1 / 31.8 / 90.7 | 1.3 / 5.7 / 25.5 | May 29, 2019, 1 p.m. |
| DenseVLAD | 7.6 / 31.2 / 91.2 | 1.0 / 4.4 / 22.7 | April 11, 2019, 5:44 p.m. |
| Localizing Visual Landmarks for Place Recognition | 7.9 / 30.0 / 85.9 | 4.1 / 15.7 / 59.1 | June 2, 2019, 10:07 p.m. |
| NetVLAD | 6.4 / 26.3 / 90.9 | 0.3 / 2.3 / 15.9 | June 2, 2019, 9:12 a.m. |
| COCL_HR | 12.7 / 60.6 / 75.4 | 0.5 / 1.6 / 2.2 | May 30, 2019, 9:20 a.m. |
| DGCNCCC | 7.2 / 28.3 / 87.6 | 0.0 / 1.1 / 13.4 | May 21, 2019, 7:38 a.m. |
| FAB-MAP | 2.7 / 11.8 / 37.3 | 0.0 / 0.0 / 0.0 | March 13, 2019, 11:50 a.m. |
| joint image segemation and depth to posenet | 0.1 / 0.8 / 31.0 | 0.0 / 0.0 / 0.0 | May 31, 2019, 11:46 a.m. |
InLoc dataset
Localization thresholds:
- All conditions: (0.25m, 10°) / (0.5m, 10°) / (1m, 10°)
| Method | duc1 | duc2 | Last updated |
|---|---|---|---|
| Visual Localization Using Sparse Semantic 3D Map | 41.4 / 59.1 / 71.2 | 38.2 / 49.6 / 58.0 | June 2, 2019, 4 a.m. |
| InLoc | 40.9 / 58.1 / 70.2 | 35.9 / 54.2 / 69.5 | April 30, 2019, 1:52 p.m. |
| DisLoc+SparsePE | 23.7 / 34.3 / 49.5 | 16.8 / 24.4 / 32.8 | May 30, 2019, 9:57 a.m. |
| DenseVLAD | 0.0 / 1.5 / 5.1 | 0.0 / 0.8 / 2.3 | April 30, 2019, 3:02 p.m. |
SILDa Weather and Time of Day dataset
Localization thresholds:
- All conditions: (0.25m, 2°) / (0.5m, 5°) / (5m, 10°)
| Method | evening | snow | night | Last updated |
|---|---|---|---|---|
| Visual Localization Using Sparse Semantic 3D Map | 29.4 / 66.5 / 93.4 | 2.2 / 11.1 / 57.4 | 20.4 / 41.8 / 69.0 | June 2, 2019, 4 a.m. |
| DenseVLAD (top-20) & D2-Net (single scale) | 24.2 / 58.0 / 76.0 | 1.0 / 6.3 / 48.1 | 23.2 / 50.2 / 79.4 | June 8, 2019, 9:18 p.m. |
| DenseVLAD (top-20) + Colmap Image Registrator | 28.2 / 56.2 / 71.2 | 3.1 / 9.4 / 36.8 | 16.3 / 33.4 / 50.7 | April 30, 2019, 12:24 p.m. |
| Localizing Visual Landmarks for Place Recognition | 0.3 / 7.1 / 41.0 | 0.0 / 0.0 / 0.9 | 0.0 / 2.7 / 56.9 | June 2, 2019, 10:06 p.m. |
| DenseVLAD | 0.2 / 6.2 / 42.1 | 0.0 / 0.0 / 0.7 | 0.0 / 2.6 / 53.9 | April 18, 2019, 12:59 p.m. |
| DGCNCCC | 0.0 / 0.0 / 0.0 | 0.0 / 0.0 / 0.0 | 0.0 / 0.0 / 0.0 | May 19, 2019, 1:16 p.m. |