A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Synthetic Training Image Dataset for Vision-Based 3D Pose Estimation of Construction Workers
Vision-based 3D pose estimation of construction workers has drawn attention for its usefulness in occupational ergonomics, safety, and productivity analysis. However, it is still challenging to develop an extensive training image dataset, which is essential for deep neural network-powered approaches, thus inhibiting the maximum potential of vision-based 3D pose estimation. To address this issue, we built a synthetic training image dataset and validated its effectiveness for 3D pose estimation. We trained and tested a state-of-the-art 3D pose estimation architecture using these synthetic images. The results show that the synthetic data-trained model can estimate 3D poses of construction workers with a Mean Per-Joint Position Error of 50.24 mm—comparable to real-data-trained model (46.5 mm). This finding indicates that synthesized construction images are effective in training a 3D pose estimation model, thus enabling the development of more accurate and scalable 3D pose estimation and alleviating the shortage of real-world construction training data.
Synthetic Training Image Dataset for Vision-Based 3D Pose Estimation of Construction Workers
Vision-based 3D pose estimation of construction workers has drawn attention for its usefulness in occupational ergonomics, safety, and productivity analysis. However, it is still challenging to develop an extensive training image dataset, which is essential for deep neural network-powered approaches, thus inhibiting the maximum potential of vision-based 3D pose estimation. To address this issue, we built a synthetic training image dataset and validated its effectiveness for 3D pose estimation. We trained and tested a state-of-the-art 3D pose estimation architecture using these synthetic images. The results show that the synthetic data-trained model can estimate 3D poses of construction workers with a Mean Per-Joint Position Error of 50.24 mm—comparable to real-data-trained model (46.5 mm). This finding indicates that synthesized construction images are effective in training a 3D pose estimation model, thus enabling the development of more accurate and scalable 3D pose estimation and alleviating the shortage of real-world construction training data.
Synthetic Training Image Dataset for Vision-Based 3D Pose Estimation of Construction Workers
Kim, Jinwoo (author) / Kim, Daeho (author) / Shah, Julianne (author) / Lee, SangHyun (author)
Construction Research Congress 2022 ; 2022 ; Arlington, Virginia
Construction Research Congress 2022 ; 254-262
2022-03-07
Conference paper
Electronic Resource
English
REALITY-ENHANCED SYNTHETIC IMAGE TRAINING DATASET FOR COMPUTER VISION CONSTRUCTION MONITORING
| TIBKAT | 2023
3D pose estimation dataset and deep learning-based ergonomic risk assessment in construction
| Elsevier | 2024