Speaker
Description
Accurate interpretation of building façades from thermal images is usually being achieved through a two-stage process, where façade layout parsing and fine-grained masonry material segmentation are carried out in sequence. Automated structural assessment, material mapping, and digital-twin generation are being enabled when these stages are completed in a reliable way, although in practice the results can depend on the dataset condition. Deep learning is generally regarded as the most effective framework for this purpose, since very high performance has been demonstrated in both pixel-level segmentation and object-level detection tasks, even if some models can behave a bit differently under varying illumination. In this article, foundational segmentation architectures such as U-Net and DeepLab are being taken into account, while state-of-the-art transformer-based models, including SegFormer , are also being considered because their capability for more global contextual reasoning has been shown.
For the detection of façade components like windows and doors, established object-detection methods—including Faster R-CNN and the YOLO family are being incorporated. When combined, these models are used for forming a practical and trainable pipeline that is compatible with MATLAB and Python environments for more comprehensive façade analysis, even if some additional tuning is usually needed. In general, accurate interpretation of building façades from thermal images is being understood as requiring these two main stages: façade layout parsing and fine-grained masonry material segmentation. Deep learning is therefore being viewed as the most robust approach, and the architectures that can be trained in MATLAB or Python are being summarised in this document together with a practical pipeline for their implementation.
