# SODA: Site Object Detection dAtaset for Deep Learning in Construction

**Authors:** Rui Duan, Hui Deng, Mao Tian, Yichuan Deng, Jiarui Lin  
**arXiv:** [2202.09554](https://arxiv.org/abs/2202.09554)  
**Date:** 2022-02-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2202.09554v1/

## Key points

- Lack of large-scale, open-source datasets for object detection in the construction industry
- Development of a new dataset called Site Object Detection dAtaset (SODA)
- Dataset contains 19,846 images with 286,201 objects categorized into 15 classes
- Dataset is diverse and voluminous, covering different site conditions, weather conditions, construction phases, angles, and perspectives
- Feasibility evaluation using two widely-adopted object detection algorithms (YOLO v3/ YOLO v4)
- Maximum mean Average Precision (mAP) achieved is 81.47%
- Total of 1,246 hours spent on data cleaning and annotation
- Data cleaning accounted for 16.4% of the time, data labeling accounted for 68%, and data checking accounted for 15.6%
- Most images in SODA have a resolution of 1920 * 1080
- Worker labels are the most common while machine and layout labels are less frequent
- Use of k-means clustering algorithm to analyze length-width ratio and range of objects in the dataset
- Statistical results showing distribution of shooting angles used to capture the images
- Contribution of a large-scale image dataset for deep learning-based object detection methods in the construction industry
- Performance benchmark set up for evaluating algorithms in this area
- Valuable resources provided for further research and development in construction site object detection

## Summary

This research paper addresses the lack of large-scale, open-source datasets for object detection in the construction industry. The authors develop a new dataset called Site Object Detection dAtaset (SODA), specifically collected and annotated for construction sites. The dataset contains 19,846 images with 286,201 objects categorized into 15 classes including workers, materials, machines, and layout. The dataset is diverse and voluminous, covering different site conditions, weather conditions, construction phases, angles, and perspectives. The authors also evaluate the feasibility of the dataset by using two widely-adopted object detection algorithms based on deep learning (YOLO v3/ YOLO v4). The evaluation shows that the dataset is suitable for typical construction scenarios and achieves a maximum mean Average Precision (mAP) of 81.47%. The research team spent a total of 1,246 hours on data cleaning and annotation. Data cleaning accounted for 16.4% of the time while data labeling accounted for 68%, and data checking accounted for 15.6%. The dataset analysis reveals that most images in SODA have a resolution of 1920 * 1080. Worker labels are the most common while machine and layout labels are less frequent. The authors also use the k-means clustering algorithm to analyze the length-width ratio and range of objects in the dataset. They provide statistical results showing the distribution of shooting angles used to capture the images. Overall, this research contributes a large-scale image dataset for deep learning-based object detection methods in the construction industry. It sets up a performance benchmark for evaluating algorithms in this area and provides valuable resources for further research and development in construction site object detection.