This document summarizes work on developing a smart parking solution using image recognition and instance segmentation. The team initially expected a pre-trained Mask R-CNN model to work well but results were poor. Through manual tuning of hyperparameters, image augmentation testing, and matching masked data to parking spots, performance improved. Key lessons were to not expect pre-trained models to transfer directly, favor simpler models, and generalize the solution to detect non-marked parking areas by identifying parking slot masks and car distances. The solution provides occupancy status classification of parking spots in images.