Two-stage edge parking occupancy system. YOLOv8s detects spaces; YOLOv8m-cls classifies each crop. On-device inference, 280B JSON output. 0.995 mAP@50 · 0.919 acc · 2291 FPS CoreML INT8 · 99.9% bandwidth saved vs 1080p stream.

How much labeled data do you actually need to deploy a parking occupancy system at a never-before-seen lot? A supervision study spanning CLIP zero-shot → ResNet-18 few-shot → full supervision on 432k parking space crops, with dataset annotation error discovery. Trained on NVIDIA A100 via IU Big Red 200.