Recycle Classifier · Jamie Kim

The motivation

Background & goal

Delivery apps made grocery and takeout ordering easier, but also increased how much packaging waste I had to sort, and figuring out what's actually recyclable isn't always obvious. The surge in packaging waste during the pandemic made this feel more urgent, which led me to explore whether machine learning could help distinguish recyclable materials from non-recyclable ones.

I used Teachable Machine to train a custom image classification model, then built a demo site with p5.js and ml5.js to run it live against a webcam feed.

Process

Training & building

1

Data & categories

500+ images · 7 classes

Trained on 6 material categories: cardboard, glass, metal, paper, plastic, and trash, plus a 7th "turn on your camera" class (a blank screen) to stop the model from misreading no-signal as an object.

2

Hyperparameter tuning

20 training runs

The "trash" category was hardest to predict; it visually overlaps with plastic and paper. Starting from defaults (50 epochs, batch size 16), I tested 20 combinations, tracking the loss-per-epoch curve each time, and landed on 65 epochs with a batch size of 512 for the most accurate results.

3

Building the demo

p5.js + ml5.js

Exported the trained model and built a live webcam classifier, learning JavaScript and p5.js along the way through Coding Train tutorials.

4

Testing

Real-world conditions

Recorded a demo testing the model against real lighting and camera distance, and adjusted the setup (darker background) to reduce misclassification.

Training results

Before & after tuning

Initial training analysis, 50 epochs, batch size 16

Initial run: 50 epochs, batch size 16

Final training analysis, 65 epochs, batch size 512

Final run: 65 epochs, batch size 512

Implementation

How the live classifier works

The demo loads the exported model, classifies each webcam frame continuously, and displays the predicted label and confidence score on screen.

// STEP 3: Get the classification!

function gotResults(error, results) {

  if (!error) {

    label = "Label: " + results[0].label;

    confidence = "Confidence: " + nf(results[0].confidence, 0, 5);

    classifier.predict(gotResults); // keep classifying

  }

}

Findings

What I learned

Training accuracy ≠ real-world accuracy. Real-time classification was much harder than the training metrics suggested: lighting, motion, and object shape all mattered.
"Trash" stayed ambiguous. Its visual overlap with plastic and paper made it the hardest category throughout.
Background mattered more than expected. A white background was frequently misread as paper, so I switched to a dark, non-white setup.
Scope had to shrink. I originally wanted a full waste-tracking dashboard, but limited JavaScript experience at the time meant the working classifier became the deliverable, and the dashboard became a longer-term goal.

Side project← Back to side projects NextSeat Timer →