♻️ Waste Sorter
Take or upload a photo of one item, and an AI model suggests which bin it belongs in. The model runs entirely in your browser: your photo never leaves your device.
Which bin?
How it works, and its limits
The model is a ResNet18 convolutional neural network, pretrained on ImageNet and fine-tuned in Lab 2 of our Deep Learning course on the TrashNet photos (by Gary Thung and Mindy Yang, MIT licence) plus a small set of real-world web photos. It was exported to the ONNX format and runs here with ONNX Runtime Web, so no server is involved.
It knows six classes: cardboard, glass, metal, paper, plastic and general trash. It works best with one item on a plain background, and it can be wrong, especially between glass and plastic. When it is less than 60% sure, it says so. Always check your local recycling rules.
Is it waste at all? A six-bin classifier always picks a bin, even for a photo of a person or a room. So a second, small model first checks whether the photo looks like a waste item. It reuses the network's general ImageNet features and was trained on our waste photos against 1,000 everyday COCO photos (people, rooms, animals, streets). On photos it had never seen, it flagged 97.5% of non-waste photos, while letting through 99.7% of TrashNet and 95.5% of real-world waste photos. On close-up photos of single objects it flagged 77%; objects made of metal or plastic, such as a French horn or a cassette player, got through about half the time, because the model judges what things look like, not what they are for. Below the threshold, the app says the photo doesn't look like waste instead of naming a bin.