Campus Image-to-GPS Localization Project

Authors: Roei Azariya Yosef, Ayala Egoz, Yair Michael Avisar
Course: Introduction to Deep Learning, Ben-Gurion University
Best Validation Error: 8.68 meters

---

Live Demo

We have deployed our model to Hugging Face Spaces. You can upload any image from the BGU campus and get its predicted location on an interactive map.

Click here to try the Live Demo

---

Project Report

A comprehensive report detailing our methodology, data curation process, architecture decisions (Trinity Network), and ablation studies.

Download / View Full Report (PDF)

---

Overview

In this repository is a deep learning solution for Image-to-GPS Regression. The goal is to predict the precise real-world location (Latitude, Longitude) of a photo taken within the university campus, utilizing only visual features.

The model

The model processes a standard $224 \times 224$ input image through a ResNet50 backbone, which we fine-tuned and injected with Spatial Dropout layers to enhance feature robustness. The extracted feature vector branches into three parallel, task-specific heads:

• Regression Head (Geometric Precision): Directly predicts the precise $(x, y)$ coordinates using MSE Loss. This head focuses on minimizing the meter-level distance error.

• Classification Head (Global Context): Classifies the image into one of 300 "Smart Zones" (generated via K-Means clustering). This provides a global context.

• Embedding Head: Extracts compact embeddings that serve as input for the Triplet Loss. We use Hard Negative Mining on these embeddings to separate confusing scenes that look alike but are far apart.

---

Data & Model Setup

Due to file size limits, the dataset and model weights are hosted on Google Drive. You must download and place them correctly for the code to run.

1. Download Model Weights (Required)

Required to run predict.py.

• File: best_model.pth
• Link: [Our Best Model]
• Action: Download and place the file in the root directory of the project.

2. Training Logs & Analysis Graphs (Optional)

For a deeper dive into the training convergence and error analysis, we provide the raw logs and the plots.

• Contains: Logs.txt, training_plot.png, localization_error_map.png
• Location: Same Google Drive folder as the model.
• Usage: Useful for verifying the results reported in the final report.

Note regarding retraining: > If you choose to train the model from scratch (by running train.py), new plots will be generated in the project folder.

3. Download Dataset

First, in the Project Folder, Create an empty folder called data. this folder should include the images folder and the CSV.

We provide two options. Option A is recommended for immediate reproduction of training.

Option A: Preprocessed Data (Ready to Train)

This version includes an images folder with the original images resized to 224*224 and contains the processed CSV (has image_name, lat, lon, utm_x, utm_y, is_night, label)

• Link: [The Processed Data Set]
• Action:
  1. Download and unzip, this folder will contain an images folder and a gt.csv.
  2. Place the images folder inside data/.
  3. Place the gt.csv file inside data/.
  4. Status: You can run python train.py immediately.

Option B: Raw Data (Submission Format)

This option follows the strict submission guidelines but requires an additional preprocessing step before training.

• Link: [The Raw Data Set]

• Structure:

  1. Download and unzip the data into the project folder.
  2. Run Preprocessing:

  python preprocess.py

  Output: This script creates a new folder named processed_data/ containing:

  • images/ (Optimized images)
  • gt.csv (Updated Ground Truth with "Smart Zone" labels)

  3. move the images folder and the gt.csv from the processed_data to the data folder
  4. And now you can run python train.py.

Final Project Structure:

Campus_GPS_Project/
├── best_model.pth          <-- Downloaded Model
├── predict.py
├── model.py
├── train.py
├── preprocess.py           <-- (Only if using Raw Data)
├── requirements.txt
└── data/
    ├── images/             <-- Images
    └── gt.csv              <-- CSV

Environment Setup

To replicate our results, please strictly follow these steps to create a clean Conda environment with the required dependencies (including utm and pillow-heif).

# 1. Create a clean environment with Python 3.9
conda create -n gps_project python=3.9 -y

# 2. Activate the environment
conda activate gps_project

# 3. Install required dependencies
pip install -r requirements.txt

---

How to Run

1. Training To train the model from scratch (ensure you followed "Data Setup Option A"):

 python train.py

2. Inference (Evaluation)

We provide a standalone function predict_gps that accepts a numpy array image and returns coordinates, exactly as required.

Example Usage:

import numpy as np
from PIL import Image
from predict import predict_gps

# 1. Load an image (Standard RGB)
img_path = "data/images/some_campus_image.jpg"
image = np.array(Image.open(img_path).convert('RGB'))

# 2. Predict Coordinates
# Returns: np.array([latitude, longitude], dtype=float32)
coords = predict_gps(image)

print(f"Predicted Location: {coords}")
# Output: [31.262345 34.803210]

---