TinyTrash

Waste classification system using computer vision, PyTorch, and embedded systems. Classifies trash into 5 categories (glass, metal, paper, plastic, others) and automatically opens the corresponding bin lid using servo-controlled mechanisms.

Complete system includes:

• ML inference on laptop (MobileNetV2)
• Wio Terminal display and control hub
• PCA9685 PWM servo driver
• 4 automated trash bins with servo-actuated lids

---

Model Training

The model uses MobileNetV2 architecture fine-tuned on a custom dataset of waste images categorized into 4 or 5 classes. Refer to tinytrash.ipynb for more information.

Categories

• 🔵 Glass - Glass bottles, jars
• ⚪ Metal - Aluminum cans, metal containers
• 🟡 Paper - Cardboard, paper packaging
• 🔴 Plastic - Plastic bottles, containers
• 🟤 Others - Anything that does not fall in the previous four categories (work in progress)

---

Dataset

Download from GitHub Releases

# Download and extract
wget https://github.com/Okikulo/TinyTrash/releases/download/v1.0/tinytrash_dataset.zip
unzip tinytrash_dataset.zip

---

Features

• Real-time waste classification using webcam
• MobileNetV2-based model for efficient inference
• Visual feedback with confidence scores
• Cross-platform support (Linux, Windows, macOS)

Requirements

• Python 3.8+
• Webcam
• PlatformIO (for Wio Terminal firmware)
• requirements.txt

Installation

# Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone repository
git clone https://github.com/Okikulo/TinyTrash.git
cd TinyTrash

# Create virtual env (optional)
uv venv

# Enter the virtual venv (if created)
source .venv/bin/activate

# Install dependencies
uv sync

Hardware Setup

Wio Terminal Firmware

1. Install PlatformIO:

pip install platformio

2. Upload firmware:

cd arduino
pio run --target upload

3. Verify upload:

• LCD should show "TinyTrash Ready"
• Serial monitor: pio device monitor

Servo Wiring

PCA9685 to Wio Terminal:

• SDA → Wio Terminal SDA pin
• SCL → Wio Terminal SCL pin
• VCC → 5V
• GND → GND

Servos to PCA9685:

• Channel 0: Paper bin servo
• Channel 1: Metal bin servo
• Channel 2: Glass bin servo
• Channel 3: Plastic bin servo

Power:

• Servos require external 5V power supply (2A+)
• Connect to PCA9685 V+ and GND terminals
• Do not power servos from Wio Terminal!

Mechanical Setup

Each trash bin requires:

• Hinged lid mechanism
• Servo horn attached to lid linkage
• 0° = closed, 90° = open
• Secure servo mounting to prevent vibration

Usage

Live Mode (inference_live.py)

Continuous real-time classification with visual overlay.

python inference_live.py

# With optional Wio Terminal display
python inference_live.py --serial --port COM3  # Windows
python inference_live.py --serial --port /dev/ttyACM0  # Linux

Features:

• Real-time classification
• FPS counter
• All class probabilities displayed
• Screenshots saved to screenshots/ folder

Capture Mode (inference_capture.py)

Capture and classify individual photos on demand.

python inference_capture.py

# With optional Wio Terminal display
python inference_capture.py --serial --port COM3

Features:

• Clean camera preview
• Capture on keypress
• Selective saving by category
• Images organized in captures/{category}/ folders
• Results in separate window

Controls

Live Mode (inference_live.py):

• p - Pause/unpause inference
• s - Save screenshot
• q - Quit

Capture Mode (inference_capture.py):

• c - Capture photo and classify
• s - Save current classification
• q - Quit
• f - Toggle fullscreen (perfect for demos!)

Command-Line Options

Both scripts support optional Wio Terminal communication:

--serial              # Enable serial communication
--port PORT           # Serial port (e.g., COM3, /dev/ttyACM0)
--baudrate BAUDRATE   # Serial baudrate (default: 115200)

Example:

python inference_capture.py --serial --port COM3 --baudrate 115200

Dataset Collection

The capture_photos.py script can help you collect data easily to add or make your own dataset. The script runs in Windows, macOS and Linux without problems.

Using the Photo Capture Tool

# Run the capture script, category argument is positional
python capture_photos.py category

# Example
python capture_photos.py glass

# For more information
python capture_photos.py --help

Controls

• SPACE - Take photo
• Q - Quit

Performance Metrics

Model Accuracy:

• 4-category model: ~92% validation accuracy
• 5-category model: ~91% validation accuracy
• Inference time: ~50-100ms

System Response:

• Classification: < 100ms
• Serial transmission: < 10ms
• Servo actuation: ~1 second (0° → 90°)
• Total time: ~1.5 seconds (capture to bin open)

Tested on:

• Laptop: NVIDIA 4060 laptop GPU / AMD Ryzen 9 CPU inference
• Wio Terminal: SAMD51 @ 120MHz
• Dataset: 1000 images across 5 categories

Extra Utilities

• converter.py to ensure all the photos are in the a compatible format (JPG,JPEG,jpg,jpeg).
• count.sh to easily count per category and the total number of photos.
• ziping.sh to easily zip all the categories.

Acknowledgments

Developed as part of Embedded Systems Design course at National Taiwan University of Science and Technology (NTUST). Special thanks to David for being an awesome and funny professor :)