A few-shot learning approach for database-free vision-based monitoring on construction sites

Abstract

This paper proposes a few-shot learning approach that can successfully learn and detect new construction objects when only a few training data are given. The proposed approach includes few-shot model design and meta-learning processes. To validate the approach, the authors conducted experiments using a popular construction benchmark dataset, AIMDataset. Even if only 20 training images were provided to a new construction object, the few-shot learning could build an object detection model with the mean Average Precision of 73.1% on average, whereas the performance of the existing supervised learning was limited to 36.5%. The results imply that the proposed approach can successfully learn and detect new types of construction objects only with few labeled images given, enabling to reduce the number of training images while maximizing the model performance. It would be then possible to save human efforts required for data labeling and enhance the practicality of vision-based construction monitoring systems.