Visual Place Recognition Project

Overview

This project implements various approaches for Visual Place Recognition (VPR) using deep learning techniques.

Experimental Branches

This repository contains nearly 20 experimental branches exploring different architectures and approaches:

Main Experimental Variants

• CCC - Core experimental implementation
• CrossSalad1, CrossSalad2, crosssalad3 - Cross-attention based variants
• PALSalad1 - Parameter-efficient learning approach
• PETLSalad1 - Parameter-efficient transfer learning variant
• mn_saladv21 - MobileNet-based variant

Additional Experiments

• add_resnet_spd - ResNet with SPD-Conv integration
• checkout - Experimental checkout branch
• easyModify - Simplified modification approach
• And more...

To view all branches:

git branch -a

Project Structure

├── dataloaders/          # Data loading utilities for various datasets
│   ├── GSVCitiesDataloader.py
│   ├── MapillaryDataset.py
│   ├── PittsburgDataset.py
│   └── val/             # Validation datasets
├── datasets/            # Dataset files and configurations
│   ├── msls_test/
│   ├── msls_val/
│   ├── Nordland/
│   ├── Pittsburgh/
│   └── SPED/
├── models/              # Model architectures
│   ├── aggregators/
│   └── backbones/
├── utils/               # Utility functions
│   ├── losses.py
│   └── validation.py
├── main.py             # Main training script
├── eval.py             # Evaluation script
└── vpr_model.py        # VPR model implementation

Installation

conda env create -f environment.yml
conda activate <env_name>

Usage

Training

python main.py

Evaluation

python eval.py

Datasets

The project supports multiple VPR datasets:

• GSV-Cities
• Mapillary Street-Level Sequences (MSLS)
• Pittsburgh
• Nordland
• SPED

License

See LICENSE file for details.