文献类型：期刊文献

英文题名：TRAINING DATA GENERATION FOR CONSTRUCTION INSTANCE SEGMENTATION FROM TARGET DOMAINS

作者：Yan, Xuzhong[1];Wang, Zeli[2]

机构：[1]Zhejiang Univ Technol, Sch Management, Dept Construct Management, Hangzhou 310023, Peoples R China;[2]East China Univ Sci & Technol, Sch Business, Shanghai 200237, Peoples R China

年份：2026

卷号：31

起止页码：632

外文期刊名：JOURNAL OF INFORMATION TECHNOLOGY IN CONSTRUCTION

收录：EI(收录号：20262220814029);WOS:【ESCI(收录号:WOS:001770184200001)】；

基金：This research was supported by the National Natural Science Foundation of China (No. 72201247 and No. 72304098) . The authors would like to acknowledge Zhejiang Construction Investment Group Co., Ltd. for visual data access.

语种：英文

外文关键词：training data generation; clustering; semi-supervised

摘要：This study addresses a persistent challenge in computer vision for construction monitoring: deep learning models trained on source-domain data often perform poorly when deployed in new target domains due to distribution shifts and limited annotations. To mitigate these issues, the research introduces TDG-CIS, a clustering-initialized semi-supervised framework designed to generate high-quality instance segmentation training data directly from unlabeled target-domain images. TDG-CIS operates in two stages. First, it employs a clustering-based mask generation strategy that uses a transformer feature backbone to extract patch-level representations and derive initial instance masks without human supervision. These masks serve as a reliable starting point for semi-supervised learning. Second, a semi-supervised instance segmentation model iteratively refines these masks and converts raw images into usable training samples. This iterative pipeline allows the model to progressively improve segmentation quality while adapting to the visual characteristics of diverse construction environments. The framework was validated on a large dataset of 50,000 images spanning more than 70 construction-related domains. Experimental results show that TDG-CIS achieves a 77.9% data utilization rate, along with 87.5% mAP and 81.1% mAR in segmentation quality. When used to scale training data for downstream instance segmentation models, TDG-CIS yields substantial performance gains: baseline models trained on automatically generated data outperform those trained on manually labeled datasets, improving mAP from 92.9% to 94.3% and mAR from 86.7% to 88.6%. Ablation studies further demonstrate that the semi-supervised refinement mechanism is key to boosting both data utilization and segmentation accuracy. Overall, the study offers a novel approach that eliminates dependence on source-domain supervision and provides a scalable pathway for producing target-domain training datasets for instance segmentation in intelligent construction applications.