SODA: A large-scale open site object detection dataset for deep learning in construction

Computer vision-based deep learning object detection algorithms have been developed sufficiently powerful to support the ability recognize various objects. Although there are currently general datasets for detection, is still a lack of large-scale, open-source dataset construction industry, which limits developments as they tend be data-hungry. Therefore, this paper develops new large-scale image specifically collected and annotated site, called Site Object Detection dAtaset (SODA), contains 15 kinds classes categorized by workers, materials, machines, layout. Firstly, more than 20,000 images were from multiple sites in different site conditions, weather phases, covered angles perspectives. After careful screening processing, 19,846 including 286,201 objects then obtained with labels accordance predefined categories. Statistical analysis shows that advantageous terms diversity volume. Further evaluation two widely-adopted based on (YOLO v3/ YOLO v4) also illustrates feasibility typical scenarios, achieving maximum mAP 81.47%. In manner, research contributes development learning-based methods industry sets up performance benchmark further corresponding area.