Rethinking Visual Geo-localization for Large-Scale Applications

The project

This work aims to implement additional features to the recent paper ”Rethinking Visual Geo-localization for Large-Scale Applications” of CVPR(2022).

You can check the full report here.

Instructions to run and test the code on Colab😄

Introduction

First of all, you should know that you can access to two Jupyter notebooks:

• The first notebook contains all the logs and tests we performed; it is accessible here.
• The second notebook contains the initial scripts you should run if you want to perform your own tests; it is accessible here.

Downloadable elements

The second notebook is ready-to-use and, by default, it downloads all datasets and models we used for our tests.

If you want to perform aimed tests with specific dataset, here is presented a table where, for each downloadable element, it is provided a brief description.

Element	Description
sf_xs	Dataset with San Francisco images
tokyo_night	Dataset with Tokyo images. It contains only night images
tokyo_xs	Dataset with Tokyo images
night_target	Dataset to perform data adaptation on night domain
logs	It contains 4 saved models: (Cosface, Sphereface, Arcface, GRL), all trained with ResNet18 as backbone
geowarp_model	It is a model that has been trained with ResNet18 as backbone and implements GeoWarp
eff2vs	It is a model that has been trained with EfficientNetV2s as backbone
eff2vs_geowarp	It is a model that has been trained with EfficientNetV2s as backbone and that implements GeoWarp
eff2vs_grl	It is a model that has been trained with EfficientNetV2s as backbone and that combines GeoWarp and GRL

Examples of commands

After you run all the scripts provided in the second notebook, you can use the following commands to run your experiments. Of course, they are just an overview, if you want to see some more example you can access to the first notebook and see.

Train CosPlace with CosFace/SphereFace/ArcFace

!python train.py --dataset_folder /content/small --groups_num 1 --epochs_num 3 --loss_function=[select one among cosface, sphereface, arcface]

Test CosPlace with CosFace/SphereFace/ArcFace

!python eval.py --dataset_folder /content/tokyo_xs/ --resume_model /content/logs/content/logs/default/[select one among trained_with_cosface, trained_with_sphereface, trained_with_arcface]/best_model.pth

Train model with GRL for domain adaptation (and with EfficientNetV2s as backbone)

!python train.py --dataset_folder /content/small --groups_num 1 --backbone efficientnet_v2_s --grl_param 0.3 --source_dir /content/small --target_dir /content/night_target

Test GRL model

!python eval.py --dataset_folder /content/tokyo-night/ --resume_model /content/logs/content/logs/default/cosplace_with_grl/best_model.pth --grl_param 0.3

Test First type of Data Augmentation

!python eval.py --dataset_folder /content/tokyo-night/ --resume_model /content/logs/content/logs/default/cosplace_with_grl/best_model.pth --grl_param 0.3 --night_test True --night_brightness 0.2

Train GeoWarp model on sf-xs and ResNet18 as backbone

!python /content/train_geowarp.py --dataset_folder /content/small --groups_num 1 --epochs_num 3

Test GeoWarp on sf-xs

!python /content/evalGeowarp.py --dataset_folder /content/small/ --resume_model /content/geowarp_model/best_model.pth

Test with multiscale

!python /content/eval.py --dataset_folder /content/tokyo-night --multi_scale --multi_scale_method=avg --select_resolution 0.526 0.588 1 1.7 1.9 --resume_model /content/logs/content/logs/default/cosplace_with_grl/best_model.pth --grl_param 0.3

Test ensembler (GeoWarp + GRL have been trained with EfficientNetV2s, NIF selected by default)

!python eval_ensemble.py --dataset_folder /content/small/ --backbone efficientnet_v2_s --grl_param 0.3 --grl_model_path /content/eff2vs_grl/eff2vs_grl.pth --geowarp_model_path /content/eff2vs_geowarp/best_model.pth

Authors

Alessio Carachino: https://github.com/CarachinoAlessio

Francesco Di Gangi: https://github.com/FDG2801