Balancing AI generalization and specialization: Multi-domain learning for universal computer vision models in construction

Abstract

While model generalization and specialization are a critical concern in computer vision, balancing them in datascarce construction settings remains challenging due to their unique nature. This paper proposes a multi-domain learning approach where a model acquires domain-generic visual knowledge from various domain datasets, while maintaining domain-specific predictabilities for each individual domain. Results show that the approach can train a more powerful model than traditional methods, regardless of training dataset size, evaluation metrics, and test domains. The model, trained on only half to one-eighth of the dataset size used in traditional methods, exhibited comparable or even superior performance while demonstrating greater robustness to challenging and diverse construction environments. These findings suggest that the approach can competitively balance model generalization and specialization, leading to improved performance across various aspects. This advance can optimize the use of given training datasets and facilitate the development of more universal computer vision models in construction.