This is the pyTorch implementation of Advanced Machine Learning course project at Politecnico di Torino. It consists in an improvement of the existing Cosplace visual-geolocalization model [ArXiv]. You can find the corresponding paper here.
The authors of this project are:
-
After downloading the dataset, to train the model you can simply run
$ python train.py --dataset_folder path/to/sf-xs -
To set groups number and epochs number:
$ python train.py --dataset_folder path/to/sf-xs --groups_num 1 --epochs_num 3 -
To change the backbone for example you can simply run
$ python train.py --dataset_folder path/to/sf-xs --backbone efficientnet_v2_s --groups_num 1 --epochs_num 3 -
To change the loss function you can simply run
$ python train.py --dataset_folder path/to/sf-xs --loss arcface --groups_num 1 --epochs_num 3 -
To perform the improved preprocessing (RandomPerspective+AdjustGamma) you can simply run
$ python train.py --dataset_folder path/to/sf-xs --preprocessing --groups_num 1 --epochs_num 3 -
Gamma preprocessing value is settable simply running
$ python train.py --dataset_folder path/to/sf-xs --preprocessing --gamma 1.5 --groups_num 1 --epochs_num 3 -
To perform the improved preprocessing (RandomPerspective+AdjustGamma) you can simply run
$ python train.py --dataset_folder path/to/sf-xs --preprocessing --groups_num 1 --epochs_num 3 -
You can perform domain adaptation with grl running
$ python train.py --dataset_folder path/to/sf-xs --groups_num 1 --epochs_num 3 --grl --dataset_root path/to/target_dataset --grl_datasets target_dataset_subfolder_name
Run $ python day_to_night_processing.py to synthetically create a night domain dataset starting from a day one.
-
You can also speed up your training with Automatic Mixed Precision
$ python train.py --dataset_folder path/to/sf-xs/processed --use_amp16
Run $ python train.py -h to have a look at all the hyperparameters that you can change. You will find all hyperparameters mentioned in the paper.
-
You can test a trained model with
$ python eval.py --dataset_folder path/to/sf-xs --backbone efficientnet_v2_s --resume_model path/to/best_model.pth -
You can test a trained model using GeoWarp as reranking method with the following script. To set custom encoder and pooling you should manually change the paths inside test.py file.
$ python eval.py --dataset_folder path/to/sf-xs --groups_num 1 --epochs_num 3 --backbone efficientnet_v2_s --resume_model path/to/best_model.pth --warping_module --num_reranked_predictions 5 -
You can test a trained model using Model soup as model ensemble method with the following script. The path folder containing .pth files of ingredients is defined inside model_soup.py files.
$ python eval.py --dataset_folder path/to/sf-xs --groups_num 1 --epochs_num 3 --model_soupe_greedy -
You can test a trained model performing FDA between images inside test queries and database with the following script.
$ python eval.py --dataset_folder path/to/sf-xs --groups_num 1 --epochs_num 3 --resume_model path/to/best_model.pth --fda --fda_weight 0.0075
Run $ python eval.py -h to have a look at all the hyperparameters that you can change. You will find all hyperparameters mentioned in the paper.
You can download plenty of trained models below.
All the trained model are downloadable at this [link].
| sf-xs (test) | Tokyo-xs | Tokyo-night |
|---|---|---|
| 53.4/66.5 | 69.5/84.8 | 50.5/72.4 |
| sf-xs (test) | Tokyo-xs | Tokyo-night | |
|---|---|---|---|
| CosPlace CosFace | 53.4/66.5 | 69.5/84.8 | 50.5/72.4 |
| CosPlace SphereFace (s=30; m=1.5) | 49.6/63.8 | 69.2/87.0 | 53.3/78.1 |
| ArcFace (s=64; m=0.5) | 47.9/61.5 | 68.9/84.1 | 53.3/72.4 |
| sf-xs (test) | Tokyo-xs | Tokyo-night |
|---|---|---|
| 67.9/77.0 | 87.0/91.7 | 80.0/87.6 |