A Generative Domain Adaptation Scheme for Swift Deployment of Parking Monitoring Systems

• Antonio Michel Ferreira dos Santos PUCPR
• Paulo Ricardo Lisboa de Almeida UFPR
• Jean Paul Barddal PUCPR
• André Gustavo Hochuli PUCPR

Resumo

Deep learning models have demonstrated remarkable accuracy in distinguishing between empty and occupied parking spaces when large amounts of annotated training data are available from the target environment. However, in real-world deployments, the major bottleneck lies in the labor-intensive annotation process required whenever a new scenario arises, or retraining is needed due to changes in the camera setup, often driven by maintenance, repositioning, or environmental conditions. This paper addresses this challenge by proposing a generative domain adaptation scheme designed to reduce annotation requirements and accelerate deployment significantly. Instead of relying on extensive labeled datasets and computationally expensive model retraining, our method synthesizes new training samples based on a small subset of instances from the target domain. In particular, by combining generative augmentation with a lightweight convolutional network for inference, our approach achieves a favorable balance between annotation cost, computational efficiency, and accuracy. These results highlight the method’s potential as a cost-effective and rapidly deployable solution for real-world parking lot monitoring. Under a cross-dataset evaluation protocol, we highlight that our approach achieves competitive accuracy (close to 97%) using as few as 256 labeled samples, thus substantially reducing human annotation effort without sacrificing classification performance.