Trash Optimizer

Context

Trash Optimizer is a two-week project developed during Le Wagon AI & Data Science Bootcamp (Nantes 2025 Q4 batch) by a team of four students:

• Paul Baudry
• Charles Poisson
• Simon Hingant
• Daria Serbichenko

Overview

Trash Optimizer addresses two key challenges in waste management: identifying the correct waste category and locating the nearest appropriate disposal facility. This application provides an end-to-end solution for waste classification and route optimization.

Key Features

• Image-Based Classification: Classifies waste items into 18 categories aligned with Nantes Metropole's waste management taxonomy
• Optimized Route Planning: Generates the shortest path to relevant collection points based on user location and waste type
• Geospatial Integration: Consolidates collection point data from multiple sources into a unified database

Technical Implementation

Machine Learning Model

• Fine-tuned EfficientNetB0 architecture for waste classification
• Trained on 3 labeled datasets comprising 18 waste categories
• Achieves 90% classification accuracy
• Extended from base model's 5 relevant categories to full 18-category taxonomy

Data Infrastructure

• Aggregated data from multiple sources: CSV files, Open Data APIs, web scraping, and geoservice APIs
• Consolidated and transformed data pipeline feeding into BigQuery
• Centralized data warehouse serving as single source of truth for collection points

Route Optimization

• Batch queries collection points from database based on waste classification
• Calculates inter-point distances using geospatial algorithms
• Generates optimized routes and provides turn-by-turn navigation

Dataset Sources

Training Datasets

• RealWaste: joebeachcapital/realwaste - UCI Machine Learning Repository (DOI)
• Recyclable and Household Waste Classification: alistairking/recyclable-and-household-waste-classification
• Custom dataset for specialized categories (batteries, light bulbs, electronics, etc.)

Collection Points Data

• Nantes Metropole Waste Management Dataset

Quick Start

Prerequisites

• Python 3.12+
• Docker (for deployment)
• API keys: Hugging Face, OpenRouteService, Google Cloud Platform

Local Development

git clone https://github.com/yourusername/trash-optimizer.git
cd trash-optimizer

# Setup individual components (see component READMEs for details)
cd training   # Model training
cd inference  # FastAPI backend
cd webapp     # Streamlit frontend

Each component has its own .env.template file - copy to .env and configure with your API keys.

Docker Deployment

cd deployment
./setup.sh
docker-compose up --build

See component READMEs for detailed setup instructions.

Architecture

graph TB
    subgraph "User Interface"
        A[Streamlit Frontend]
    end

    subgraph "Backend Services"
        B[FastAPI Inference Service]
        C[BigQuery Database]
    end

    subgraph "External Services"
        D[OpenRouteService API]
        E[Hugging Face Hub]
    end

    subgraph "Training Pipeline"
        F[Dataset Builder]
        G[Model Training]
        H[Kaggle Datasets]
    end

    A -->|Image Upload| B
    B -->|Download Model| E
    A -->|Query Collection Points| C
    A -->|Route Optimization| D
    B -->|Classification Result| A

    H -->|Download| F
    F -->|Processed Dataset| G
    G -->|Upload Trained Model| E

    style A fill:#e1f5ff
    style B fill:#ffe1e1
    style C fill:#e1ffe1
    style F fill:#fff4e1
    style G fill:#fff4e1

Loading

Data Flow:

1. User uploads waste image → Inference API classifies using model from Hugging Face Hub
2. Classification result + user location → Query BigQuery for relevant collection points
3. Collection points → OpenRouteService calculates optimal route → Display to user

Sub components documentation

• Training Module: Fine-tune classification models on custom waste datasets
• Inference Backend: FastAPI service for waste classification
• Webapp Frontend: Streamlit interface for classification and route optimization
• Deployment: Docker containerization and cloud deployment guides

Model Repository

Models are hosted on Hugging Face Hub: https://huggingface.co/cpoisson/trash-optimizer-models

Docker Deployment

For a local or production deployment, use Docker to run both the inference backend and webapp in a single container:

cd deployment
./setup.sh              # Initial setup
docker-compose up --build  # Build and run

See deployment/README.md for complete deployment documentation.

Known Limitations

• Collection point data is currently specific to Nantes Metropole area
• Model trained primarily on household waste items

Contributing

This project was developed as an educational bootcamp project at Le Wagon. We welcome feedback, bug reports, and suggestions via issues.

Acknowledgments

• Le Wagon - AI & Data Science Bootcamp (November 2025)
• Dataset Contributors: RealWaste (UCI), Kaggle community datasets
• Nantes Metropole - Open data initiative for collection points
• External Services: Hugging Face Hub, OpenRouteService, Google BigQuery

License

This project is licensed under the MIT License