This code is built upon the official repo for Neurocomputing 2022 paper GSV-Cities: Toward Appropriate Supervised Visual Place Recognition

[ArXiv]

• The GSV dataset can be dowloaded from Dataset. Make a folder corresponding to each city inside datasets/gsv-cities/images/City.
• Upload the metada for each city in datasets/gsv-cities/Dataframes.
• The Nordland Dataset, SPED, and Pittsburg datasets can be downloaded from here.
• The MSLS data set can be downloaded from the here Mapillary Street-level Sequences Dataset.
• Unzip the dataset files into their corresponding folders inside datasets.

• Training can be run from main.py. The code has comments for using datasets for different cities, different aggregators and benchmark datasets.
• The files for different aggregators are present in models/aggregators. The aggregator fields and configurations are commented and can be changed in VPRmodel.
• The validation data set can be changed through GSVCitiesDataModule(val_set_names = [...]).
• The fast_dev_run field can be used before training to check if the code is in order.

The code to run the model from some previous checkpoint is as follows

from main import VPRModel

model = VPRModel(backbone_arch='resnet50', 
                 layers_to_crop=[],
                 agg_arch='...',
                 agg_config={...},
                )

state_dict = torch.load('./LOGS/resnet50...')
model.load_state_dict(state_dict)
model.eval()