MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering

Environment Setup

Our experiments were conducted using three NVIDIA RTX 3090 GPUs. Please follow the steps below to configure the environment using the provided requirements.txt file:

conda create -n MutualVPR python=3.8
conda activate MutualVPR
pip install -r requirements.txt

---

Dataset Preparation

We use the SF-XL dataset in our experiments. It can be downloaded by following the instructions in the official gmberton/CosPlace repository (CVPR 2022 paper: "Rethinking Visual Geo-localization for Large-Scale Applications").

The base dataset used in the experiments is derived from processed dataset folder, and the cropping strategy requires cropping the panoramic images. Panoramic images are included in the raw dataset folder.

Optional: Multi-angle Cropping

To apply a multi-view cropping strategy, utility functions are provided in datasets/dataset_utils.py. You can manually set the following parameters:

• Starting angle for cropping
• Saving path
• Cropping step size, etc.

After preparing the data, modify the initialize() function in train_dataset.py to include the newly cropped data.

Training

Before training, make sure to modify the dataset paths in the code to match your local setup.

Run the training script with:

python train.py
--train_set_folder /your/path/train
--val_set_folder /your/path/val
--test_set_folder /your/path/test

🔗 Model Weights

Our pretrained weight file can be downloaded from the following link:

👉 Download from Google Drive

Testing

To evaluate SF-XL-testv1 and SF-XL-occlusion, you also need to download queries-v1 and queries-occlusion from CosPlace. Both of them use the same database.

The datasets/ directory includes test splits for the following benchmarks:

• SF-XL
• Tokyo 24/7
• MSLS
• Pitts250k
• Pitts30k

To test on any of these datasets, select your checkpoints and simply run:

python test.py

Acknowledgement

We would like to thank the authors of CosPlace and EigenPlaces for their great work and generously providing source codes, which inspired our work and helped us a lot in the implementation.