Bolivian Architecture Classification System

Computer Vision Model for Identifying Traditional Bolivian Architectural Styles

Python AWS S3 Stable Diffusion Roboflow Machine Learning 98.1% Accuracy
View Live Project See Results

Project Overview

An end-to-end computer vision system for architectural heritage classification

Challenge

Develop an automated system to identify and classify traditional Bolivian architectural styles, including indigenous Cholet buildings, Adobe houses, Standard residential homes, and modern Buildings - preserving cultural heritage through AI.

Solution

Built a custom object detection model using Roboflow 3.0, integrating AWS S3 for scalable data storage and Stable Diffusion for synthetic data generation to augment the training dataset.

Impact

Achieved 98.1% precision and 96.5% recall across 721 annotated images, enabling automated architectural analysis for urban planning, cultural preservation, and real estate applications.

Technology Stack

Modern cloud-native tools for scalable ML development

AWS S3

Cloud storage for dataset management

Stable Diffusion

Synthetic image generation

Roboflow

ML platform for training & deployment

Python + Boto3

Data pipeline automation

Key Metrics

Performance results demonstrating model effectiveness

98.1%
Precision
96.5%
Recall
721
Images Trained
4
Architecture Classes

System Architecture

End-to-end data pipeline and model workflow

Data Sources

Internet + S3 + Stable Diffusion

→
Annotation

Manual labeling in Roboflow

→
Training

Roboflow 3.0 Fast Model

→
Deployment

Cloud-hosted inference

Development Process

Systematic approach from data collection to model deployment

1

Data Collection & Storage

Gathered 721 images from multiple sources including public repositories and internet sources. Stored dataset in AWS S3 bucket for scalable access and integrated with Roboflow for automated batch uploading.

2

Synthetic Data Generation

Implemented Stable Diffusion in SageMaker Studio Lab to generate additional training images, addressing class imbalance and expanding dataset diversity for more robust model training.

3

Image Annotation

Manually labeled all images using Roboflow's annotation tools (Polygon, Smart Polygon, Bounding Box) to define 4 architectural classes: Cholet, Adobe, Standard, and Building.

4

Model Training & Evaluation

Trained Roboflow 3.0 Fast model with 100 epochs, 16 batch size, and 80/10/10 train/validation/test split. Achieved 98.1% precision and 96.5% recall through iterative optimization.

Model Predictions

Real-world classification examples showcasing model accuracy

Building classification

Edificio (Building)

Confidence: 70-90%

Modern high-rise building successfully detected with precise bounding box localization.

Cholet classification

Cholet

Confidence: 87%

Traditional Andean architecture recognized with distinctive color patterns and structural elements.

Adobe classification

Adobe House

Confidence: 88%

Traditional adobe construction identified by characteristic earthen materials and rural setting.

Standard house classification

Estándar (Standard House)

Confidence: 88%

Contemporary residential architecture classified with high confidence and accurate localization.

Try the Model with Your Own Images

Key Technical Learnings

Insights gained throughout the project lifecycle

Class Imbalance

The Edificio class had 396 annotations vs 197 for Estándar, requiring careful monitoring for potential bias. Implemented data augmentation and synthetic generation to address this challenge.

Cloud Integration

Successfully orchestrated AWS S3, Roboflow API, and Stable Diffusion across multiple cloud platforms, demonstrating proficiency in distributed ML workflows.

Model Selection

Chose Roboflow 3.0 Fast model balancing inference speed with accuracy requirements. Optimal for real-time architectural analysis applications.