Scan to open OopsBin

·

OopsBin

Multiclass waste classification system powered by deep learning. Upload an image or use your camera and OopsBin tells you which bin it belongs in.

7 categories | 93.3% accuracy | 4 models | 3 languages (EN / ES / KO)

---

Models

Four models were trained and compared. ResNet50 was selected as the default for production due to its best overall accuracy.

Model	Backbone	Classes	Test Accuracy	Status
ResNet50 (R)	ResNet50	7	93.3%	Default
EfficientNetB0 Model 1 (J)	EfficientNetB0	7	Available	Active
EfficientNetB0 Model 2 (R)	EfficientNetB0	7	Available	Active
MobileNetV2 (N)	MobileNetV2	10	Available	Disabled (different classes)

All models can be selected from the Classify tab in the web interface.

---

Quick Start

Prerequisites

• Python 3.10+
• uv (recommended) or pip
• Docker (optional)

1. Clone the repository

git clone https://github.com/Bootcamp-IA-P6/project4_Team_5.git
cd project4_Team_5

2. Set up environment variables

Create a .env file in the project root:

SUPABASE_URL=supabase_url
SUPABASE_KEY=supabase_anon_key

3. Run the application

Option A: uv (recommended)

# Install uv (if not installed)
# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

uv sync

# macOS / Linux
make dev

# Windows (make not available)
uv run python app/main.py

Open http://localhost:8080

Option B: pip

python -m venv venv

# macOS / Linux
source venv/bin/activate

# Windows CMD
venv\Scripts\activate

# Windows PowerShell
venv\Scripts\Activate.ps1

pip install -r requirements.txt
python app/main.py

Open http://localhost:8080

Option C: Docker

# macOS / Linux
make docker

# Windows / any OS
docker compose up --build

Open http://localhost:8080

Make commands (macOS / Linux)

Command	Description
make dev	Run locally with uv
make docker	Build and run with Docker
make docker-down	Stop Docker containers
make docker-rebuild	Rebuild Docker from scratch
make clean	Remove all Docker images, volumes and cache

---

Project Structure

project4_Team_5/
├── app/                          # Web application
│   ├── main.py                   # Flask backend + prediction API
│   ├── templates/                # HTML pages
│   │   ├── base.html             # Base layout, theme, i18n
│   │   ├── index.html            # Main app (Home, Classify, Metrics, About, Contact)
│   │   └── 404.html              # Error page
│   └── static/
│       ├── css/style.css         # Styles (light/dark, accessibility)
│       ├── js/main.js            # Frontend logic, camera, feedback
│       ├── img/                  # Logo, mascot, characters
│       └── music/                # Background audio
│
├── models/                       # Trained models
│   └── versions/                 # ResNet50, EfficientNetB0 x2, MobileNetV2
│
├── notebooks/                    # Training & data preparation notebooks
│
├── config/                       # Configuration files
│
├── data/                         # Data management
│
├── tests/                        # Unit tests
│
├── Dockerfile                    # Container image
├── docker-compose.yml            # Container orchestration
├── Makefile                      # Dev shortcuts
├── requirements.txt              # Python dependencies (pip)
├── pyproject.toml                # Python dependencies (uv)
└── .env                          # Environment variables (not in git)

---

Features

• Image classification via upload or drag & drop
• Camera classification with real-time capture
• Multi-model selection — choose between ResNet50, EfficientNetB0 (x2), or MobileNetV2
• User feedback system connected to Supabase (PostgreSQL)
• Metrics dashboard with confusion matrix, F1 scores, learning curves, and error analysis
• Multilingual support (English, Spanish, Korean)
• Accessibility — dark/light mode, colorblind modes, font size control, high contrast
• Dockerized for easy deployment
• Custom 404 error page

---

Tech Stack

Layer	Technology
Backend	Flask, Python
ML	TensorFlow / Keras (ResNet50, EfficientNetB0, MobileNetV2)
Data	NumPy, Pandas, Pillow, Matplotlib
Frontend	HTML, CSS, JavaScript
Database	Supabase (PostgreSQL)
Tooling	uv, Make, Jupyter, Git LFS
Container	Docker, Docker Compose
Training	Google Colab / Kaggle (GPU T4)

---

Datasets

Three public datasets from Kaggle were combined and unified:

Dataset	Original Classes	Images
Garbage Classification	6	2,527
Garbage Classification (12 classes)	12	15,515
RealWaste	9	4,752
Total after unification	7	16,756

Final classes: cardboard | glass | metal | organic | paper | plastic | trash

Split: Train 11,036 (80%) | Validation 1,376 (10%) | Test 1,385 (10%)

---

API Endpoints

Method	Endpoint	Description
GET	/	Main application
GET	/models	List available models
POST	/predict	Classify an image (supports model selection)
POST	/feedback	Submit prediction feedback
GET	/feedback/stats	Get feedback statistics

---

Training Summary

The default model (ResNet50) was trained in 3 phases using transfer learning:

Phase	Epochs	Learning Rate	What trains	Val Accuracy
Phase 1 — Head only	20	1e-3	Classification head (526K params)	91.5%
Phase 2 — Fine-tuning	15	1e-5	Last 30 layers + head	94.2%
Phase 3 — Extended	4 (early stop)	5e-6	Last 30 layers + head	94.3%

Final results: 93.3% test accuracy | 2.6% overfitting (below 5% threshold)

---

Team

Naizabeth Bermudez

Jonathan Brasales

Raul Machaca Luna

Roberto Molero

Recycle right. The planet thanks you.