All work
Computer Vision2024
Evaluating Instance Segmentation for Waste Detection
A research project evaluating whether instance segmentation models can realistically run on resource-constrained sorting or inspection hardware — not just how accurate they are.
Overview
A computer vision project comparing SOLOv2, Mask R-CNN and U-Net for waste segmentation on the TACO dataset, with a multi-axis evaluation framework — accuracy, deployment efficiency and compute cost — built to compare them fairly under resource-limited deployment constraints.
What I built
- Designed a controlled comparison holding backbones, dataset, and training regime constant across all three architectures.
- Built a three-axis evaluation framework (mAP, overlap-threshold efficiency, FLOPs/parameter count) instead of ranking on accuracy alone.
- Diagnosed and fixed a broken, unmaintained data/tooling pipeline (duplicate annotations, missing files, incompatible library versions) end to end.
- Wrote up findings responsibly — citing a peer-reviewed benchmark rather than presenting an unfinished in-house result as final.