README.md

---
license: mit
task_categories:
- image-to-text
tags:
- gps
- geolocation
- computer-vision
- regression
- campus
pretty_name: Image2GPS Penn Campus
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data_car/01_Split_Dataset/train-*
  - split: validation
    path: data_car/01_Split_Dataset/validation-*
  - split: test
    path: data_car/01_Split_Dataset/test-*
dataset_info:
  features:
  - name: image
    dtype: image
  - name: Latitude
    dtype: float64
  - name: Longitude
    dtype: float64
  splits:
  - name: train
    num_bytes: 1879686949
    num_examples: 525
  - name: validation
    num_bytes: 106413976
    num_examples: 29
  - name: test
    num_bytes: 106913215
    num_examples: 30
  download_size: 2093075615
  dataset_size: 2093014140
---

# 🗺️ Image2GPS — Penn Campus Dataset

Geotagged image dataset for predicting GPS coordinates from photos taken on the University of Pennsylvania campus.

## 📋 Overview

| | Detail |
|---|---|
| 🎯 **Task** | Image → GPS regression (latitude, longitude) |
| 📍 **Region** | Penn campus: 33rd & Walnut → 34th & Spruce |
| 📷 **Sources** | Human-height (\~1.5m) & car-height (\~0.15m) |
| 📊 **Metric** | Average Haversine distance (meters) ↓ |

## 📁 Structure

```
data_human/
└── 01_Split_Dataset/
    ├── train/          # ~1595 images + metadata.csv
    ├── validation/     # ~199 images + metadata.csv
    └── test/           # ~200 images + metadata.csv

data_car/              # Car-height images (exploratory)
```

Each `metadata.csv` contains:

| Column | Description |
|--------|-------------|
| `file_name` | Image filename |
| `Latitude` | GPS latitude in decimal degrees |
| `Longitude` | GPS longitude in decimal degrees |

## 🚀 Quick Start

```python
from datasets import load_dataset

# Load human-height data
dataset = load_dataset("Wu52F/Image2GPS_dataset", data_dir="data_human/01_Split_Dataset")

train = dataset["train"]
print(train[0])  # {'image': <PIL>, 'Latitude': 39.952, 'Longitude': -75.193}
```

## 📷 Collection Protocol

- Photos taken along walkways on Penn campus
- 8 photos per location (rotating 360°)
- Phone held upright, no zoom
- GPS extracted from EXIF metadata
- HEIC images converted to JPEG with EXIF preserved

## 👥 Team

CIS 5190 Applied Machine Learning — Spring 2026  
Team 15: Tao Wu, Yuchen Xu