UAV-Collected Multi-Class Instance Segmentation Dataset for Building Facades Defects

Current building inspections are mainly relied on human, making the process time-consuming and labor-intensive. Therefore, the development of deep-learning-based automated inspection systems has become a growing trend. However, the lack of a comprehensive training dataset is the majority challenge in developing deep-learning models. To bridge this gap, we introduce CUBIT-CW, a human-annotated dataset containing 10871 high-resolution images collected via drone, with multiscale ranging from <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$800 \times 600$</tex> to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$8000 \times 6000$</tex> pixels. Our CUBIT-CW covers 9 distinct defect categories, which are line cracks, map cracks, rust stains, delamination, spalling, cement loss, seepage, damp patches, efflorescence and mold. It also works for instance segmentation, allowing for the models to have a detailed understanding of defect boundaries and features. To evaluate its effectiveness, we test CUBIT-CW by 6 state-of-the-art(SOTA) computer vision models. The corresponding experimental results illustrate that our dataset enhances the robustness and generalizability of the models, making them to inspect more complex defects on building surfaces with greater accuracy.