This repository contains the code for our papers Learning Condition Invariant Features for Retrieval-Based Localization from 1M Images and Geometrically Mappable Image Features. The corresponding models and training/testing image lists can be downloaded here.
This code was tested using TensorFlow 1.10.0 and Python 3.5.6.
It uses the following git repositories as dependencies:
The training data can be downloaded using:
| Name | Model |
|---|---|
| I: Triplet | triplet_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| I: Quadruplet | quadruplet_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| I: Lazy triplet | lazy_triplet_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| I: Lazy quadruplet | lazy_quadruplet_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| I: Triplet + distance | sum_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| I: Triplet + Huber dist. | h_sum_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| II: Triplet | triplet_5e-6_full-10-25_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| II: Quadruplet | quadruplet_5e-6_full-10-25_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| II: Lazy triplet | lazy_triplet_5e-6_full-10-25_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| II: Lazy quadruplet | lazy_quadruplet_5e-6_full-10-25_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| II: Triplet + Huber dist.~ | h_sum_5e-6_full-10-25_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| III: Triplet | triplet_xy_000 |
| III: Quadruplet | quadruplet_xy_000 |
| III: Lazy triplet | lazy_triplet_xy_000 |
| III: Lazy quadruplet | lazy_quadruplet_xy_000 |
| III: Triplet + distance | distance_triplet_xy_000 |
| III: Triplet + Huber dist. | huber_distance_triplet_xy_000 |
| III: Triplet + HP | evil_triplet_hp_001 |
| III: Quadruplet + HP | evil_quadruplet_hp_001 |
| III: Volume without HP | ha0_loresidual_det_muTrue_renone_vl64_pca_eccv_000 |
| *III: Volume without HP** | ha0_loresidual_det_muTrue_renone_vl0_pca_eccv_001 |
| III: Volume | residual_det_0.0_eccv_000 |
| **III: Volume*** | residual_det_512_none_0_fc_eccv_000 |
| IV: Triplet | pittsnetvlad |
| V: Off-the-shelf | offtheshelf |
| Name | Oxford Model |
|---|---|
| Off-the-shelf | offtheshelf |
| Triplet + dist. | distance_triplet_xy_000 |
| Triplet + Huber dist. | huber_distance_triplet_xy_000 |
| Lazy quad. + dist. | distance_lazy_quadruplet_xy_000 |
| Lazy quad. + Huber dist. | huber_distance_lazy_quadruplet_xy_000 |
| Dist. | distance_xy_000 |
| Huber | huber_distance_xy_000 |
| Triplet | triplet_xy_000 |
| Quadruplet | quadruplet_xy_000 |
| Lazy triplet | lazy_triplet_xy_000 |
| Lazy quadruplet | lazy_quadruplet_xy_000 |
| Name | COLD Model |
|---|---|
| Off-the-shelf | offtheshelf |
| Triplet + dist. | sum_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| Triplet + Huber dist. | h_sum_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| Lazy quad. + dist. | nlq_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| Lazy quad. + Huber dist. | hlq_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| Dist. | simple_dist_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| Huber | huber_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| Triplet | triplet_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| Quadruplet | quadruplet_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| Lazy triplet | lazy_triplet_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |
| Lazy quadruplet | lazy_quadruplet_5e-6_all_conditions_angle_1-4_cu_LRD0.9-5_noPCA_lam0.5_me0 |