Geolocalization Project

Overview

This project focuses on visual place recognition and geolocalization using deep learning techniques. It implements and experiments with various models, loss functions, and training strategies to improve the accuracy of place recognition across different datasets and introduces DeMixVPR framework.

Table of Contents

1. Installation
2. Project Structure
3. Usage
4. Experiments
5. Datasets
6. Models
7. Evaluation

Installation

To set up the project environment, follow these steps:

1. Clone the repository:

2. Create a virtual environment (optional but recommended):

   conda create --name geolocalization --file requirements.txt
   conda activate geolocalization
   

Project Structure

The project is organized as follows:

• experiments.ipynb: Main Jupyter notebook for running experiments
• models/: Directory containing model architectures
• datasets/: Custom dataset classes for training and testing
• utils/: Utility functions, including custom learning rate schedulers
• train.py: Training script
• evaluation.py: Evaluation script
• weights/: Directory to store trained model weights
• data/: Directory to store RGB images and depth maps datasets.

Usage

To run experiments and train models, use the experiments.ipynb notebook. This notebook contains various experimental setups and configurations for training and evaluating models.

Experiments

The project includes several experiments with different configurations. Some of them are:

1. Average Pooling + Contrastive Loss
2. GEM Pooling + Contrastive Loss
3. Batch Hard Miner + Contrastive Loss
4. DistanceWeightedMiner + SupConLoss
5. PairMargin Miner + FastAP Loss
6. MultiSimilarity Loss

Each experiment can be run by setting the TRAIN variable to True and executing the corresponding cell in the notebook.

Datasets

The project uses the following datasets:

• GSVCities-XS Dataset: For training
• SF_XS Dataset: For validation and testing
• Tokyo-XS Dataset: For testing

Dataset loading and preprocessing are handled in the datasets/ directory.

Models

The main model architecture is defined in the models/helper.py file. It uses a truncated ResNet18 as the backbone with various pooling options (average, GEM).

Evaluation

Model evaluation is performed using the eval_model function from evaluation.py. It calculates metrics such as Recall@1 and Recall@5 for the trained models on different test sets.

Weights and Dataset

You can download the pre-trained weights and dataset from the following link: https://drive.google.com/drive/folders/1C7qqH9qhxOm5dKoo6csMS1ldJrPJQzEo?usp=sharing

Report

Report of the project can be found in the repository as report.pdf.