Streamlit app comparing VLM-only, YOLO-only, and hybrid approaches for parking-lot occupancy detection on the same image. Built with YOLO26 (PKLot) + NVIDIA LocateAnything-3B.

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 493,654 parking space crops, with dataset annotation error discovery. Trained on NVIDIA A100 via IU Big Red 200.

Parking space occupancy classifier using classical computer vision (GLCM, Sobel, Prewitt, Harris corners). Per-park calibration with majority voting. 89–92% accuracy on PKLot dataset. No deep learning. Interactive Tkinter demo included.