Waste stream classification
Instance segmentation of conveyor imagery for recycling AI: about two years of monthly batches for a confidential EU client.
Problem
Recycling AI must separate paper, cardboard, Tetra and other classes on conveyor imagery, so sorting systems can recognise each material as it passes.
Data and task
Instance segmentation on 512 by 512 pixel conveyor images, including hybrid classes for items that do not fall cleanly into one material.
Tools
Supervisely, used for segmentation and review throughout the programme.
Challenges
Overlapping materials and keeping class boundaries consistent at production volume: every mask has to follow the same rules late in the programme as it did at the start.
QA
Sampled error review, with about 1% error found in an audited sample.
Delivery scale
About two years of monthly batches and millions of objects, with roughly 16,800 objects per batch in audited samples, for a confidential EU client.
Result and next step
A long-running segmentation programme delivered in monthly batches. This is the latest account of our waste-management work in the Netherlands.
Related industry: Recycling and environment. Related service: Computer Vision.