Image Geolocation Prediction Using k-NN and CLIP Embeddings

Overview

This project predicts the geographic coordinates (latitude and longitude) of images using the k-Nearest Neighbors (k-NN) algorithm. The dataset comprises geo-tagged images from Flickr, and we leverage OpenAI's CLIP embeddings for image feature analysis.

Key Features

• Predicts image locations using k-NN.
• Employs PCA for dimensionality reduction.
• Optimizes k-NN with grid search for the best k-value.
• Evaluates performance using Mean Displacement Error (MDE).

Requirements

The following Python libraries are required:

• matplotlib
• numpy
• scikit-learn

Installation

1. Clone the repository
2. Navigate to the project directory
3. Install dependencies

Usage

1. Place the dataset file im2spain_data.npz in the project directory.
2. Run the main script:

Output

• Visualizations of image locations and features.
• Mean Displacement Error for different k-values.
• Comparison of k-NN with Linear Regression.

Notes

Ensure that the required Python libraries are installed before running the code. For any issues, please contact the project maintainer.