A Practical Image Augmentation Method for Construction Safety Using Object Range Expansion Synthesis

Jaemin Kim, Ingook Wang, Jung-Ho Yu, Seulki LeeBuildings 2025 · DOI

TL;DR

This paper introduces Object Range Expansion Synthesis (ORES), a practical and efficient method for generating synthetic data for construction safety applications. ORES overcomes the challenges of collecting real-world data in hazardous environments by synthesizing realistic images from existing object masks and backgrounds, enabling better object recognition models with less manual effort.

What Problem Does This Paper Address?

Construction sites are dynamic, hazardous, and data-scarce environments. Training AI models for tasks like object recognition, crucial for safety applications (e.g., detecting scaffolding, workers), requires large, high-quality datasets. However, collecting such data is difficult due to safety concerns, logistical challenges, and high costs. Existing data augmentation techniques often lack the realism needed for construction environments, and 3D modeling approaches are computationally expensive and require specialized expertise. This gap hinders the widespread adoption of AI in construction safety.

Methods

The core of this paper is the Object Range Expansion Synthesis (ORES) method. It's a lightweight, non-generative approach that works by:

  1. Object Separation: Extracting object masks from existing images (both real and from open datasets).
  2. Background Inpainting: Filling in the areas where objects were removed from the background.
  3. Augmentation & Synthesis: Applying classical augmentations (like color adjustments) to the extracted objects and then synthesizing them into various background scenes.

A key aspect is that ORES preserves object scale and ground alignment, ensuring realistic placement. It leverages existing open datasets and a small amount of labeled data, minimizing the need for extensive manual labeling or complex 3D modeling. The process is automated, allowing for rapid generation of diverse datasets.

Datasets

  • Training Data: 30,000 synthetic images generated using the ORES method.
  • Real Data for Training/Validation: 4,000 real construction site images were used as a base for generating synthetic data.
  • Testing Data: 1,500 real construction site images were used for model evaluation.
  • Open Datasets: Used for background scenes and object masks. Specific datasets are not detailed but are referred to as "open datasets" and "AI-Hub (a publicly available dataset platform for AI research in Korea)."

Key Insights and Ideas

  • Practicality over Novelty: The authors prioritize a method that is directly applicable and cost-effective for real-world construction safety problems, even if the underlying techniques aren't entirely novel.
  • Leveraging Existing Resources: ORES effectively utilizes readily available open datasets and a small amount of real data, significantly reducing the burden of data collection and labeling.
  • Addressing Linear Objects: The method is specifically highlighted for its ability to handle linear objects like scaffolding and ladders, which are often underrepresented and crucial for fall-related accident detection.
  • Scalability: The approach is designed for large-scale dataset generation, making it suitable for complex and dynamic environments.

Key Results

  • High Precision: The object recognition model trained with ORES achieved a mean average precision (mAP50) of 98.74% on real-world construction site images.
  • Moderate Recall: The model achieved a recall of 54.55%. While this indicates room for improvement, the high precision demonstrates the practical value of the synthetic data.
  • Efficiency: 30,000 synthetic images were generated in just 30 minutes using 4,000 real images as a base, showcasing the speed and efficiency of the method.
  • Improved Accuracy with Epochs: The model's performance (mAP) showed improvement as the number of training epochs increased, indicating effective learning from the synthetic data.

Limitations and Future Work

  • Low Recall: The authors acknowledge that the recall is relatively low, suggesting that some objects, particularly scaffolding, were missed. This is attributed to the visual similarity between objects and backgrounds, hindering feature extraction.
  • Simplified Viewpoint Alignment: The current implementation uses only three categories for viewpoint alignment (frontal, side, oblique), which is insufficient for accurate visual alignment, leading to unnatural synthesized images in some cases.
  • Future Work Suggestions:
    • Object-wise Analysis: Conduct detailed precision and recall analysis for different object types to identify specific weaknesses.
  • Complex Visual Conditions: Incorporate advanced augmentation techniques to simulate occlusion, partial visibility, lighting variations, and viewpoint diversity.
  • Hybrid Training: Explore combining real and synthetic data in a hybrid strategy to bridge the domain gap and improve overall robustness and recall.
  • Field Experiments: Conduct real-world experiments (e.g., crane automation, drone inspections) to validate the practical feasibility and performance of ORES.
  • Broader Object Scope: Expand the training data to include more object types and high-risk scenarios to enhance generalizability.
  • Background Matching: Develop mechanisms to selectively pair backgrounds with similar color and environmental features to improve realism.

Summary updated

Abstract

This study aims to propose a practical and realistic synthetic data generation method for object recognition in hazardous and data-scarce environments, such as construction sites. Artificial intelligence (AI) applications in such dynamic domains require domain-specific datasets, yet collecting real-world data can be challenging due to safety concerns, logistical constraints, and high labor costs. To address these limitations, we introduce object range expansion synthesis (ORES), a lightweight and non-generative method for generating synthetic image data by inserting real object masks into varied background scenes using open datasets. ORES synthesizes new scenes, while preserving scale and ground alignment, enabling controllable and realistic data augmentation. A dataset of 30,000 synthetic images was created using the proposed method and used to train an object recognition model. When tested on real-world construction site images, the model achieved a mean average precision at IoU 0.50 (mAP50) of 98.74% and a recall of 54.55%. While recall indicates room for improvement, the high precision highlights the practical value of synthetic data in enhancing model performance without requiring extensive field data collection. This research contributes a scalable approach to data generation in safety-critical and data-deficient environments, reducing dependence on direct data acquisition, while maintaining model efficacy. It provides a foundation for accelerating the deployment of AI technologies in high-risk industries by overcoming data bottlenecks and supporting real-world applications through practical synthetic augmentation.
BibTeX
@inproceedings{kim2025practical,
  author = {Jaemin Kim and Ingook Wang and Jung-Ho Yu and Seulki Lee},
  title = {{A Practical Image Augmentation Method for Construction Safety Using Object Range Expansion Synthesis}},
  booktitle = {Buildings},
  year = {2025},
  doi = {10.3390/buildings15091447},
}
Figure