LoQI-VPR

VPR for Low Quality Images via Knowledge Distillation

Implementation for the IROS 2025 submission "Distillation Improves Visual Place Recognition for Low Quality Images".

Setup

1. Clone the repository with submodules: $ git clone --recurse-submodules https://github.com/ai4ce/LoQI-VPR.git
2. Install dependencies: $ conda env create -f environment.yml
3. Download GSV-Cities dataset from Kaggle and the Pitts250k dataset for validation
4. Download VPR testing datasets using VPR Datasets Downloader

Running Experiments

trainer.yaml and test_trained_model.yaml from configs contains the configurations for running distillation and testing VPR methods respectively.

Training: src/trainer_gsv-cities.py distills VPR models enabled in configurations using the enabled loss functions.

Testing: src/dataset/testing_data.py precomputes global descriptors for the specified VPR models and datasets selected in configurations. src/calculate_recall.py records recall rates to tensorboard log files and a Google Sheet.