Segmentation Dataset for Reinforced Concrete Construction

TL;DR

This work presents ConRebSeg, a publicly released dataset of 14,805 RGB images with four-class segmentation (Exposed rebars, Persons, Cars, Trucks) targeting autonomous inspection of reinforced concrete. It benchmarks three baselines (YOLOv8L-seg, DeepLabV3, U-Net) and analyzes how data availability and labeling styles affect performance, revealing that pre-trained YOLOv8L-seg achieves the strongest validation performance (up to around $0.59$ mIOU) while withholdings and label inconsistencies have limited impact. An error analysis via TIDE shows missed detections as the primary failure mode, underscoring the need for more data and robust localization. The study also demonstrates practical deployment considerations on embedded hardware (Jetson Xavier NX) with FP16 offering a favorable balance between accuracy and speed, and advocates for broader open data sharing in construction to address data scarcity and improve model generalization.

Abstract

This paper provides a dataset of 14,805 RGB images with segmentation labels for autonomous robotic inspection of reinforced concrete defects. Baselines for the YOLOv8L-seg, DeepLabV3, and U-Net segmentation models are established. Labelling inconsistencies are addressed statistically, and their influence on model performance is analyzed. An error identification tool is employed to examine the error modes of the models. The paper demonstrates that YOLOv8L-seg performs best, achieving a validation mIOU score of up to 0.59. Label inconsistencies were found to have a negligible effect on model performance, while the inclusion of more data improved the performance. False negatives were identified as the primary failure mode. The results highlight the importance of data availability for the performance of deep learning-based models. The lack of publicly available data is identified as a significant contributor to false negatives. To address this, the paper advocates for an increased open-source approach within the construction community.

Figures (13)

• Figure 1: Visual examples provided as annotation guidelines
• Figure 2: Class distribution within each split
• Figure 3: Histogram of normalized label centroids in the train (upper row), val (middle row), and test split (bottom row) of the ConRebSeg dataset, for the mask bounding boxes (left) and for the mask itself (right)
• Figure 4: File system structure of the dataset
• Figure 5: Sequence of successive frames with their labels (in red) showing an inconsistency in labeling style. The left and right frame share the same annotation style, the center frame has a different style.