Hierarchical GeoCLIP

The project implements GeoCLIP, a CLIP-inspired image geolocation model. Hierarchical GeoCLIP is an improved inference method that leverages hierarchical feature clustering at multiple geographical scales. By organizing the GPS gallery into a tree structure, we drastically reduce the search space and achieve comparable performance while being ~100x more efficient.

Instructions to run

1. Clone the repository and go to the project folder.

git clone https://github.com/ramanakshay/hierarchical-geoclip
cd hierarchical-geoclip

2. Install dependencies from requirements file. Make sure to create a virtual environment before running this command.

pip install -r requirements.txt

4. Download dataset from scripts inside the data folder. Run main.py to train the model

python main.py

Requirements

• pytorch (An open source deep learning platform)
• hydra (A framework for configuring complex applications)
• sklearn (For clustering algorithms)
• datasets (Hugging Face Datasets)

Folder Structure

├──  model              - this folder contains all code (models, networks) of the agent
│   ├── classifier.py
│   ├── network.py
│   └── layers.py
│
│
├── data                - this folder contains code relevant to the data
│   ├── preprocessing.py         - run this script to download the data   
│   └── dataset.py             - dataset classes

│
│
├── algorithm             - this folder contains different algorithms of your project
│   └── train.py
│
│
├──  config
│    └── config.yaml  - YAML config file for project
│
│
├──  clustering            - this folder contains code for clustering the GPS gallery
│    └── cluster.py        - build JSON file from list of GPS coordinates and pre-trained model
│
│
└── main.py           - entry point of the project

License

This project is licensed under the MIT License. See LICENSE for more details.