Material tracking for Paper Recycling

This project involved automation of multiple processes in a recycling facility using computer vision and robotics. In the first phase, a vision based system was designed and deployed on forklifts to intelligently classify different classes of paper that look very similar to each other, but require very different processes for recycling.

The main challenges in this task were a large number of classes (80+) that visually look very similar and skewed distributions (70%+ crates were a single type of paper, while many classes contributed to less than 0.1% of the crates). Another challenge was unavailability of data from the recycling facility, and therefore, having to rely on datasets from a different distribution. These challenges were addressed using methods such as domain adaptation, contrastive learning, and hierarchical classification.

Role in Project: Leading the computer vision research