Pranay004/high-density-object-segmentation

Instance segmentation and object counting in crowded scenes. A fine-tuned YOLOv8s-seg model runs entirely in the browser via ONNX Runtime Web — no server, no upload.

Compares the detector alone against a density-aware second pass gated on Canny edge density. Measured on 50 held-out COCO val2017 images: count accuracy rises from 52% to 64%, MAE falls from 4.28 to 3.30.

Source, ablation and 18-page report: https://github.com/Zakuroooo/high-density-object-segmentation