Adaptive object detection for construction sites via parameter-efficient fine-tuning with LoRA

Construction sites present diverse and evolving visual conditions that challenge the generalizability of pre-trained object detection models. This study proposes a parameter-efficient fine-tuning approach based on Low-Rank Adaptation to enable adaptive object detection tailored to site-specific conditions. A general model was trained on a large-scale dataset and fine-tuned using both the proposed method and full fine-tuning across three real-world construction projects. Despite utilizing only 12% of the trainable parameters, the proposed approach achieved comparable or superior detection accuracy with 10% reduced training time and 30% lower GPU memory consumption. These results highlight its effectiveness in adapting object detection models to site-specific conditions under resource constraints. Furthermore, the approach can be extended with semisupervised learning to support scalable adaptation in construction environments.