Moving-Feature-Driven Label Propagation for Training Data Generation from Target Domains

Deep learning models, due to their high sensitivity to training data distributions, may suffer from performance reduction when applied to construction sites different from the source domain where the training data originated. Although various technologies to improve the generalization capability of deep learning models, such as transfer learning, synthetic data generation, few-shot learning, and domain adaptation, collecting training data from a new target domain is generally inevitable to attain a desirable monitoring performance. To overcome the laborious processes of data re-collection and annotation from a new target domain, this paper presents a self-training strategy to generate training data for construction object detection. The proposed method produces target domain training data by: (1) employing optical flow estimation to detect moving objects from the target domain, (2) leveraging self-training to propagate existing labels to unlabeled data (moving objects), and (3) utilizing image inpainting and copy-paste augmentation to augment target domain-specific training data. Experimental results from four different scenes demonstrate the efficacy of the proposed method in boosting the performance of object detectors within new target domains. The findings of this study will advance technologies of improving the generalization of deep learning models, thereby facilitating automated monitoring systems for the construction domain.