overhead-person-yolo11s

YOLO11s fine-tuned for overhead / ceiling-camera person detection. Part of an on-device (browser WebGPU) cooking-robot safety system.

• Base: yolo11s.pt (COCO-pretrained), full fine-tuning
• Data: overhead-person (Roboflow Universe, CC BY 4.0), re-split 80/10/10
• Single class: person

Results (test split)

metric	stock yolo11s (before)	fine-tuned (after)
recall	0.442	0.980
precision	0.627	0.969
mAP@50	0.506	0.991
mAP@50-95	0.160	0.752

Files

• best.pt — PyTorch weights (ultralytics)
• best.onnx — ONNX opset 12, for onnxruntime-web / WebGPU on-device inference

Usage

from ultralytics import YOLO
YOLO("best.pt").predict("image.jpg", conf=0.25)

Code & write-up: https://github.com/chanubc/overhead-person-yolo11