PaperlayerView exact revision on arXiv

Source-linked AI summary

SODA: Site Object Detection dAtaset for Deep Learning in Construction

Rui Duan, Hui Deng, Mao Tian, Yichuan Deng, Jiarui Lin

arXiv:2202.09554v1cs.CV

TL;DR

Construction object detection lacks a large, open, domain-specific dataset despite deep-learning algorithms being data-hungry. This paper constructs and releases SODA, a 15-class construction-site dataset assembled from diverse field imagery, and reports benchmark evaluations showing its feasibility for construction scenarios.

  • Problem

    Construction object detection lacks a large, open dataset, while deep-learning algorithms require substantial data and existing construction datasets are limited.

  • Method

    The paper builds SODA by collecting, cleaning, annotating, and benchmarking construction-site imagery across workers, materials, machines, and layout.

  • Results

    SODA contains 19,846 images, 286,201 objects, and 15 categories, and was evaluated with mainstream one-stage object-detection algorithms.

  • Takeaways & Limitations

    SODA provides a publicly available construction-site dataset and benchmark for subsequent deep-learning object-detection research.

  • Takeaways & Limitations

    SODA has fewer categories and images than major deep-learning datasets, and its object-level bounding-box annotations cannot train pixel-level segmentation algorithms.

Abstract

from arXiv · show

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.

1. Introduction

Construction management remains labor-intensive, while deep-learning object detection needs large, domain-specific datasets. The study addresses this gap by building and publicly providing SODA for construction-site detection research.

  • Manual management and intervention make construction-site management difficult and inefficient.
  • Deep-learning object detection is data-hungry, and construction imagery is difficult to collect and annotate because construction is highly specialized.
  • SODA contains 19,846 construction-site images, 286,201 objects, and 15 manually annotated categories.
  • The images were collected from more than 20,000 visual records using monocular cameras, UAVs, and hook-visualization equipment, then processed and annotated by trained civil-engineering students.
  • SODA provides 15 construction-site classes across workers, materials, machines, and layout, with publicly released images and annotations.

2. Related Work

Existing computer-vision datasets include broad general collections and smaller construction-specific datasets. Construction research still needs broader coverage of materials and site layouts alongside workers and machines.

  • General image datasets support computer-vision development but often lack categories specific to construction sites.
  • Existing construction datasets commonly focus on equipment, guardrails, PPE, or other narrow applications.
  • Construction datasets have concentrated on workers, PPE, and some machines, while materials and site layouts remain less represented.
  • Object detection predicts both category and location information, with one-stage methods generally faster but slightly less accurate than two-stage methods.

3. Methodology

SODA was constructed through category selection, field image acquisition, data cleaning, and annotation. The methodology combines multiple viewpoints and explicit cleaning procedures to cover construction-site objects and conditions.

  • 3. Methodology: The dataset-building process comprises category selection, image acquisition, image cleaning, and image annotation.
  • 3.1 Category Selection: SODA includes 15 target classes organized into workers, materials, machines, and layout categories.
  • 3.2 Data acquisition: 21,863 images were collected from real construction sites across construction stages using UAVs, handheld cameras, and monitoring video.
  • 3.2 Data acquisition: Multiple shooting methods and viewpoints were combined because site confusion, blind spots, and occlusion make positive-sample collection difficult.
  • 3.3 Data cleaning: Data cleaning removes duplicates, ambiguities, non-targets, and privacy-sensitive content before annotation.

1. Remove duplicates

Duplicate images were manually removed to ensure meaningful visual differences among dataset images.

  • Repetitive images from repeated handheld shots and video-frame extraction were manually removed.

2. Remove ambiguities

Blurred images were removed because they are difficult to annotate and can impair deep-learning model training.

  • Blurred images were deleted before or during annotation to avoid annotation difficulty and degraded model training.

3. Remove non-targets

Images without dataset targets were manually removed because they were irrelevant to the research.

  • Images lacking dataset targets were manually removed because they were irrelevant to the research.

4. Privacy protection

The dataset underwent privacy processing, annotation controls, error correction, and extensive quality checking before release.

  • Privacy protection: Privacy processing covered company logos and information as well as faces, which were manually blurred by students.The processing addressed commercial-secret, property-rights, and ethical concerns.
  • Annotation standards: Annotation followed VOC conventions and required boxes to frame targets while minimizing irrelevant background.Occluded or inconveniently annotatable targets were omitted, and nearby similar targets received separate boxes.
  • Annotation workflow: 35 trained students used labelImg to annotate 15 categories and generate XML files containing image and coordinate information.The authors monitored progress and inspected the resulting annotation documents with another expert.
  • Quality control: Annotation errors included spelling, pluralization, unknown labels, and shortcut-key mistakes, all corrected in the released dataset.The reported shortcut-key errors included 48 ‘w’ and 2 ‘wwww’ labels.

4. Statistics of the dataset

SODA contains 19,846 images and 286,201 objects across diverse construction-site views, with coverage spanning workers, materials, machines, and layout.

  • Dataset scale: 19,846 images contain 286,201 objects, with 86% of images sized 1920 * 1080.Workers have the largest label count, while machine and layout labels are less numerous.
  • Dataset scale: Each class contains more than 1000 targets, and each category contains more than 20,000 targets.
  • Object statistics: The dataset uses k-means clustering to analyze object length-width ratios and ranges for detector sample requirements.The clustering results are presented in Figure 11.
  • Viewpoint diversity: Images were captured from handheld short- and long-range views, UAVs, and tower-crane hook visualization equipment.These perspectives were used to achieve full coverage, with the shooting-angle distribution shown in Figure 12.
  • Dataset comparison: Compared with existing construction datasets, SODA has the largest numbers of objects and categories and covers workers, materials, machines, and layout.It also introduces hook-visualization-equipment imagery, described as neglected in previous studies.

5. Experiments on the dataset

The experiments evaluate SODA with YOLO v3 and YOLO v4 using object-detection metrics and training curves. YOLO v4 achieves higher mAP and detection speed, while performance varies across object categories and scene complexity.

  • Experimental setup: SODA was evaluated with the one-stage object-detection algorithms YOLO v3 and YOLO v4 to provide a benchmark for subsequent research.The evaluation used the dataset’s training and test splits and 100 training epochs with freezing and thawing stages.
  • Evaluation metrics: mAP combines Precision and Recall to evaluate classification and localization across object-detection classes.Intersection over union measures overlap between detection and ground-truth boxes before calculating classification outcomes used in evaluation.
  • Training results: The training curves indicate that both deep learning algorithms fit the SODA dataset well.The paper cautions that loss values across algorithms are not necessarily comparable because they use different loss functions.
  • Overall performance: 81.47% mAP was achieved by YOLO v4, compared with 71.22% mAP for YOLO v3.These are the reported overall mAP values on SODA.
  • Overall performance: 31.94 FPS was achieved by YOLO v4, compared with 25.06 FPS for YOLO v3.Category-level performance also differed: material detection had higher mAP than worker detection, while the best and worst AP categories varied by model.
  • Practical verification: YOLO v4 identified more categories and achieved higher recognition accuracy than YOLO v3 in complex scenes, while their performance was similar in simple scenes.The trained detectors were applied to actual complex construction-site images, where the identified classes were reported as correctly classified.

6. Conclusion and future work

SODA provides a 19,846-image, 286,201-object construction-site dataset spanning 15 categories and establishes benchmarks for object detection. The authors identify limits in category and data volume, annotation granularity, model coverage, and manual annotation, motivating future expansion.

  • Dataset and benchmarks: 19,846 images and 286,201 objects across 15 categories comprise SODA, divided into training and test sets at a 9:1 ratio.The dataset uses VOC format and covers common construction-site objects.
  • Dataset and benchmarks: The dataset establishes training results and benchmarks using mainstream one-stage object detection algorithms for subsequent construction-industry research.The paper frames these benchmarks as a basis for selecting detectors in later applications.
  • Dataset and benchmarks: SODA covers workers, materials, machines, and layouts across construction stages, viewpoints, and times of day, supporting broad construction-site coverage.The paper describes near, far, vertical, UAV, and monitoring perspectives.
  • Limitations and future work: The authors recommend adding categories and data because SODA remains smaller than datasets in the broader deep learning community.This limitation persists despite SODA exceeding other construction-industry datasets in category and image count.
  • Limitations and future work: SODA provides object-level bounding-box annotations rather than pixel-level labels, so it supports object detection but not object segmentation training.Pixel-level annotation is identified as future work.
  • Limitations and future work: The study trains only two classical deep learning models and relies on manual annotation by engineering-management students, leaving broader model and annotation methods for future work.The authors mention additional models, crowdsourcing, and automated annotation as possible extensions.

Credit authorship contribution

The authors’ contributions span methodology, conceptualization, validation, resources, analysis, writing, and funding acquisition across the listed researchers.

  • Rui Duan led methodology and contributed formal analysis, investigation, and the original draft.
  • Hui Deng contributed resources, validation, review and editing, and funding acquisition.
  • Mao Tian contributed methodology, validation, and review and editing.
  • Yichuan Deng led conceptualization and contributed methodology, review and editing, and funding acquisition.
  • Jiarui Lin contributed conceptualization, resources, validation, and review and editing.
Loading 2202.09554v1…