ConViD — Concrete Visual Defect Dataset

This dataset represents a large and diverse set of real-world photos showing different types of concrete surface defects, such as spalling honeycombing voids, and cracks. The images were taken with regular consumer-grade mobile phones outside under different weather and light conditions in Pune, India. This makes the dataset very relevant to real-world situations and very useful for example in the field. The data is organized in a way that makes it easy to develop and compare automated diagnostic systems based on deep learning, especially convolutional neural networks (CNNs). Unlike most existing datasets which focus on identifying fractures only, this dataset enables the multi-class classification of different types of defects that not only look different but also require different repair methods as per the engineering standards like ACI.

One of the major characteristics of this dataset is its multi-layered complexity. Several types of defects, especially spalling, cracks, and honeycombing, visually look very similar, which makes the separation between them a difficult task. This not only poses a complex problem for computational models but also serves as a strong benchmark for evaluating fine-grained image categorization, feature representation learning, and domain adaptation methods. First round of testing show significant overlapping characteristics among different classes which is a further proof for it being a difficult experimental setting. Thanks to its real acquisition conditions and variability, this dataset is perfect for improving models that can be run on commonly used hardware, for instance, smartphones, and it is these kinds of situations where the gap between research and practical structural assessment is getting close