Answers are grounded in this paper and cite their sources.
Driven by the digital transformation of the construction industry, multimodal vision-language models have been widely applied to construction scene perception. However, mainstream methods are constrained by poor domain semantic adaptation, low cross-modal matching accuracy, and high manual annotation costs. This paper proposes a BLIP-2 multimodal scene understanding method with a self-accumulated semantic mechanism. A semantic progressive domain fine-tuning strategy is adopted based on indoor construction image-text datasets to improve the model’s descriptive capability for construction-specific objects and behaviors. Experimental results on an indoor renovation corridor show that the optimized model achieves BLEU, ROUGE-L and METEOR scores of 0.14, 0.44 and 0.35, respectively, with a key scene element recall rate of 0.89, substantially outperforming the original BLIP-2 model. Validated for narrow indoor construction scenarios, the proposed lightweight model serves as a front-end semantic perception module for digital twin systems. It supports fine-grained object recognition, scene semantic abstraction, real-time environmental perception, construction behavior analysis and risk early warning in intelligent construction. The framework delivers stable and quasi-real-time image parsing in the test scenario, while its robustness and cross-scenario scalability require further experimental verification. Compared with traditional manual-dependent and static visual detection schemes, this method reduces annotation biases and provides an innovative approach for intelligent construction site monitoring.
No takes yet. Share an insight, caveat, or question.
- et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: