Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Training a Visual Scene Understanding Model Only with Synthetic Construction Images
While the use of deep neural networks (DNN) for computer vision is increasing in the construction domain, the shortage of training data sets prevents such models from achieving their maximum potential. To address this issue, we investigate the potential of using synthetic data for vision model development. Specifically, we synthesize construction images and train a DNN model only with the synthetic data. We then evaluate the performance of the synthetic data-trained model on a worker detection task, and the results demonstrate the great potential of synthetic images: 97.3% of mean average precision. Given the benefits of synthetic data—it is possible to automatically create an unlimited number of images without manual labeling—this finding is promising. Moreover, this approach can be readily applied to other computer vision tasks, without requiring the manual labeling. This finding will enable the creation of more accurate and scalable DNN models for construction applications.
Training a Visual Scene Understanding Model Only with Synthetic Construction Images
While the use of deep neural networks (DNN) for computer vision is increasing in the construction domain, the shortage of training data sets prevents such models from achieving their maximum potential. To address this issue, we investigate the potential of using synthetic data for vision model development. Specifically, we synthesize construction images and train a DNN model only with the synthetic data. We then evaluate the performance of the synthetic data-trained model on a worker detection task, and the results demonstrate the great potential of synthetic images: 97.3% of mean average precision. Given the benefits of synthetic data—it is possible to automatically create an unlimited number of images without manual labeling—this finding is promising. Moreover, this approach can be readily applied to other computer vision tasks, without requiring the manual labeling. This finding will enable the creation of more accurate and scalable DNN models for construction applications.
Training a Visual Scene Understanding Model Only with Synthetic Construction Images
Kim, Jinwoo (Autor:in) / Kim, Daeho (Autor:in) / Shah, Julianne (Autor:in) / Lee, SangHyun (Autor:in)
ASCE International Conference on Computing in Civil Engineering 2021 ; 2022 ; Orlando, Florida
Computing in Civil Engineering 2021 ; 221-229
24.05.2022
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Training a Visual Scene Understanding Model Only with Synthetic Construction Images
| British Library Conference Proceedings | 2021
Automated annotation for visual recognition of construction resources using synthetic images
| Online Contents | 2016
Automated annotation for visual recognition of construction resources using synthetic images
| British Library Online Contents | 2016