Train with augmented dataset

If you want to apply augmentation to the datasets with translation models refer image-to-image translation/README.md and after augmenting run

$ python3 train.py --is_aug --dataset_folder path/to/sf-xs/processed

Train

After downloading the SF-xs dataset, simply run

$ python3 train.py --dataset_folder path/to/sf-xs/processed

the script automatically splits SF-XS in CosPlace Groups, and saves the resulting object in the folder cache. By default training is performed with a ResNet-18 with descriptors dimensionality 512 is used, which fits in less than 4GB of VRAM.

To change the backbone or the output descriptors dimensionality simply run

$ python3 train.py --dataset_folder path/to/sf-xs/processed --backbone resnet50 --fc_output_dim 128

You can also speed up your training with Automatic Mixed Precision (note that all results/statistics from the paper did not use AMP)

$ python3 train.py --dataset_folder path/to/sf-xs/processed --use_amp16

Run $ python3 train.py -h to have a look at all the hyperparameters that you can change. You will find all hyperparameters mentioned in the paper.

Reproducibility

Results from the paper are fully reproducible, and we followed deep learning's best practices (average over multiple runs for the main results, validation and hyperparameter search on the val set).

Test

You can test a trained model as such

$ python3 eval.py --dataset_folder path/to/sf-xs/processed --backbone resnet50 --fc_output_dim 128 --resume_model path/to/best_model.pth

You can download plenty of trained models below.

SF-N Dataset

You can download the SF-N dataset using this link.