Garbage Classifier

Take a photo of a piece of rubbish and sort it into one of six materials: cardboard, glass, metal, paper, plastic, or trash. The data is the Kaggle garbage classification set. The question we wanted to answer was how far the classical methods get you before a convolutional network is worth the trouble.

Three models, same problem

Model	Test accuracy
Support vector machine	54%
Random forest	56%
Convolutional network	74%

The two classical models land within two points of each other, which tells you something on its own. All the extra flexibility in the random forest bought almost nothing, because what was holding both of them back was the features, not the classifier. The CNN works out its own features and picks up eighteen points over the better of the two.

The failures are more interesting than the score

The per-class numbers say a lot more than 74% does. Paper is basically solved at 0.97 recall. Glass is the worst class in the set at 0.50 recall, and most of what it loses goes to metal, which is why metal has the worst precision at 0.56.

That confusion isn’t random. Shot under the same lighting, a crushed glass fragment and a crumpled can look almost the same, and with a few hundred training images per class there isn’t much for the model to separate them with.

The repo also has the exploratory analysis and a small Streamlit app that runs the trained CNN on a photo you upload.