Chiara Romanengo

Window detection on building façades is a relevant task for several applications in urban mapping, 3D city modeling, building characterization and semantic annotation, e.g., for energy performance assessment.However, façade windows remain difficult to identify reliably in real urban scenes. Image-based methods can achieve good detection performance, but they may still suffer from occlusions or produce false positives on urban objects with a window-like appearance, such as vehicle windshields. Conversely, methods operating directly on point clouds or mesh suffer from poor acquisition of transparent surfaces, which typically result in sparse or incomplete 3D data.In this paper, we propose a multimodal framework combining a learning-based window detection from spherical imagery with semantic filtering on a 3D point cloud reconstructed from the same images. The key idea is to exploit the semantic and geometric consistency of the 3D scene to validate image-based detections and reject candidates that do not belong to building façades, exploiting the complementary strengths of 2D visual information and 3D semantic-geometric context. The proposed framework is evaluated on a street-level urban dataset acquired in Catania. Experimental results show that the proposed 3D validation stage significantly improves the reliability of the detections, reducing false positives by more than 50\% and increasing precision from 0.58 to 0.70, while preserving competitive overall detection performance. In addition to the proposed pipeline, we introduce a new manually annotated dataset derived from spherical street-level acquisitions in Catania.

Advancements in 3D data acquisition technologies have significantly increased the availability of point clouds (PCs) in urban areas, leading to a growing demand for the accurate and efficient construction of urban models at various levels of geometric and semantic detail. Numerous methods have been developed to address this problem in computer vision, photogrammetry, remote sensing and computer graphics, often tailored to specific applications. In this context, the proposed survey focuses on approaches that generate 3D urban models starting from city PCs produced by various acquisition devices. The strategies used to create 3D models are grouped according to whether they use templates, basic surface primitives, hybrid approaches or the identification of linear primitives (such as edges and contours). We analyse the different processes applied to the input PC and the type of output representation created, with more details on the reconstruction of buildings, as these are key urban objects in most of the techniques proposed in the literature which are crucial to test and train the developed methods.

We propose CurveML, a benchmark for evaluating and comparing methods for the classification and identification of plane curves represented as point sets. The dataset is composed of 520k curves, of which 280k are generated from specific families characterised by distinctive shapes, and 240k are obtained from Bézier or composite Bézier curves. The dataset was generated starting from the parametric equations of the selected curves making it easily extensible. It is split into training, validation, and test sets to make it usable by learning-based methods, and it contains curves perturbed with different kinds of point set artefacts. To evaluate the detection of curves in point sets, our benchmark includes various metrics with particular care on what concerns the classification and approximation accuracy. Finally, we provide a comprehensive set of accompanying demonstrations, showcasing curve classification, and parameter regression tasks using both ResNet-based and PointNet-based networks. These demonstrations encompass 14 experiments, with each network type comprising 7 runs: 1 for classification and 6 for regression of the 6 defining parameters of plane curves. The corresponding Jupyter notebooks with training procedures, evaluations, and pre-trained models are also included for a thorough understanding of the methodologies employed.