BinSight

Johns Hopkins places recycling bins with several sorting categories across campus, but students frequently put items in the wrong bin, either from not knowing which category an item belongs to or from the time it takes to sort correctly in passing. BinSight addresses that by using computer vision to identify an item from a camera image and assign it to one of four bins, trash, recycling, compost, or paper. The full pipeline runs locally on a $166 Raspberry Pi build.

How it works

A Raspberry Pi Camera 3 feeds a live video stream into the frontend, where the user holds the item inside a bounding box. The frame selection runs a pixel-wise stability check: frames are sampled every 500ms, and once the image stays within a pixel-wise difference of 25 for 1500ms, the region inside the box is cropped and sent to the backend. A Flask service runs an EfficientNet-B2 classifier, trained in PyTorch and fine-tuned on a composite of the MJU-Waste, TrashNet, CompostNet, and Fruits and Vegetables Image Recognition datasets, re-split by hand to match Johns Hopkins recycling rules into 1,982 compost, 1,570 trash, 590 recycling, and 571 paper images, with image augmentation to even out the differences between the naturalistic and web-scraped sources. The predicted bin is then shown back on the screen for the user.

Why it's interesting

In a user study with 15 students at Brody Learning Commons, students found BinSight easier and more straightforward to use than sorting items themselves, and rated it more helpful than the standard bin labels, 3.73 against 2.33 on a 5-point scale. The classifier reached 0.927 average accuracy across the four classes, so the ease of use did not trade off against putting items in the right bin.

The code is on GitHub at github.com/sameraslan/BinSight.