[Paper Review] SODA: Site Object Detection dAtaset for Deep Learning in Construction

Rui Duan, Hui Deng+3 more|arXiv (Cornell University)|Feb 19, 2022

Infrastructure Maintenance and Monitoring4 citations

TL;DR

This paper introduces SODA, a large-scale, open-source dataset of 19,846 construction site images with 286,201 annotated objects across 15 categories (workers, materials, machines, layout). Collected under diverse conditions and evaluated with YOLOv3/v4, SODA achieves a maximum mAP of 81.47%, establishing a benchmark for deep learning-based object detection in construction.

ABSTRACT

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

Motivation & Objective

• To address the lack of large-scale, open-source datasets tailored for construction site object detection.
• To collect and annotate diverse, real-world construction site images across varying conditions, phases, and perspectives.
• To provide a benchmark dataset that supports the development and evaluation of deep learning-based object detection algorithms in construction.
• To improve the performance and generalization of object detection models in complex, dynamic construction environments.

Proposed method

• Collected over 20,000 images from multiple active construction sites under varying weather, lighting, and construction phases.
• Screened and processed images to ensure quality and consistency, resulting in 19,846 final images for the dataset.
• Annotated all objects using 15 predefined categories: workers, materials, machines, and layout elements.
• Employed bounding box annotations with class labels to support object detection training and evaluation.
• Evaluated the dataset using two state-of-the-art YOLO-based models (YOLOv3 and YOLOv4) to validate its usability.
• Conducted statistical analysis to demonstrate dataset diversity and volume, confirming its representativeness for real-world construction scenarios.

Experimental results

Research questions

• RQ1 Can a large-scale, diverse dataset of construction site images support effective training of deep learning object detection models?
• RQ2 How does the performance of standard YOLO-based models generalize on construction-specific object detection tasks using the SODA dataset?
• RQ3 To what extent does the diversity of site conditions, object categories, and viewpoints in SODA enhance model robustness and accuracy?
• RQ4 Does the SODA dataset provide a reliable benchmark for future algorithm development and evaluation in construction site monitoring?

Key findings

• The SODA dataset comprises 19,846 high-quality images with 286,201 annotated objects across 15 construction-relevant categories.
• The dataset exhibits strong diversity in terms of site conditions, weather, construction phases, and viewing angles, enhancing model generalization potential.
• Evaluation on YOLOv4 achieved a maximum mean average precision (mAP) of 81.47%, demonstrating the dataset’s suitability for training high-performing models.
• Statistical analysis confirms the dataset’s balanced distribution across object categories and robust representation of real-world construction variability.
• The dataset sets a new performance benchmark for construction site object detection, enabling standardized evaluation of future algorithms.
• The open-sourcing of SODA facilitates reproducible research and accelerates innovation in AI-driven construction site monitoring.