Synthetic Images Impact in Parking Spot Classification using Neural Networks

Urban growth and the increasing number of vehicles have intensifiedthe demand for efficient parking management solutions. In thiscontext, machine learning-based image monitoring systems havegained prominence due to their low cost and ease of installationcompared to traditional methods, such as physical sensors. Thesesystems achieve an average accuracy of 95% in cross-validationscenarios using well-known datasets like PKLot and CNRPark-EXT.However, despite the availability of extensive datasets, challengesremain regarding the accessibility and diversity of training data.This is especially critical when aiming to improve the accuracy ofgeneralist models or specialize them for specific scenarios, whereeach application requires a substantial effort to collect, segment, andlabel new images for optimal performance. This study proposes theuse of synthetic images, generated with the Unity 5 graphics engineand the Unity Perception package, to complement or replace realdata in training parking classification models. A synthetic imagegeneration protocol was developed to reduce costs compared to thecollection, segmentation, and labeling of real images. The imagesgenerated through this protocol are referred to as low-fidelity dueto their lower quality and reduced capacity to simulate specific environments.Using MobileNetV3 and transfer learning, experimentswere conducted in three scenarios: total replacement of real data,supplementation of diverse datasets, and specialization for specificscenarios. The results showed that synthetic images could improvemodel generalization by up to 2% in datasets with limited real data(e.g., CNRPark-EXT). However, synthetic images alone could notfully replace real data due to their limited fidelity in replicatingreal-world conditions, reinforcing the need for combinations withreal data or more realistic synthetic data for better results.