GeoGuessr Assistant

Type: Computer Vision System

Primary Language: Python

Published: 2026

Accuracy:

74.8%

City-level predictions

Repository: GitHub

GeoGuessr Assistant is a computer vision system that analyzes Google Street View images and predicts geographic locations. The system achieves 74.8% accuracy at the city level by analyzing road signs, text, and other visual elements through machine learning.

Overview
Capabilities
Methodology
Architecture
Getting Started
Future Development
See Also

Overview

This project was developed as a final-year computer science dissertation (3rd year Licence). The system represents an attempt to solve the geolocation problem—determining where a photograph was taken based purely on visual information—through intelligent analysis of road signage and textual elements within Street View images.

Rather than making direct predictions, the system employs a novel approach based on elimination logic, progressively narrowing the search space by identifying regional and national characteristics visible in the imagery.

Key Capabilities

Road Sign Detection: Fine-tuned YOLOv8m model with three classification categories (MUTCD, Vienna Convention, Ambiguous)
Optical Character Recognition: EasyOCR with integrated language detection
Probabilistic Reasoning: Custom elimination and probability logic
Geographic Integration: OpenStreetMap integration via Overpass and Nominatim APIs
Local Execution: Runs entirely offline without cloud dependencies
Performance: Fast inference suitable for real-time predictions

Methodology

Detection Pipeline

The system processes Street View images through the following pipeline:

Image acquisition and preprocessing
Road sign detection using fine-tuned YOLOv8m
Text extraction via EasyOCR with language identification
Geographic reasoning through elimination
OpenStreetMap validation
Location prediction with confidence scoring

Sign Classification

The YOLO model classifies detected signs into three categories to infer geographic region:

MUTCD Signs: Indicates North American locations (United States, Canada)
Vienna Convention Signs: Indicates European and international locations
Ambiguous Signs: Requires additional contextual analysis

Architecture

Component Overview

Component	Technology	Purpose
Object Detection	YOLOv8m (fine-tuned)	Road sign detection
Text Recognition	EasyOCR	Extract text from images
Language Detection	langdetect	Identify text language
Geographic Data	OpenStreetMap (Nominatim, Overpass)	Location validation and enrichment

Getting Started

For detailed installation and usage instructions, see the project repository on GitHub.

Demo

The following video demonstrates the system in operation:

Future Development

Planned enhancements to the system include:

Vegetation and climate zone detection
Building architecture recognition and classification
Vehicle model identification for temporal context
Additional visual cues (utility poles, road markings, etc.)
Improved web interface for accessibility
Multi-image analysis for increased accuracy