The paper is focused on an ongoing project funded by the Emilia-Romagna region and aimed at the creation of a new workflow finalizing digital data from integrated survey towards an “adaptive” Building Information Modeling (BIM). The project AIM-eBIM—Adapted Information Management for existing Buildings Information Modeling—brings together regional research laboratories and companies to pursue industrial research topics towards a greater deployment of digital tools. Digital surveying has triggered huge potential for innovation, but also generated new challenges in managing and using large amounts of data, often left unused. The quantity of surveyed data used to document built or Cultural Heritage often does not correspond to the quality or reliability of information. Moreover, parametric modeling of existing heritage through BIM is becoming as pervasive as it is necessary, considering regulatory trends. However, these tools can be ineffective from the point of view of users (professionals, companies) who must deal with such complexity. The challenge is to bring discretization (and simplification) processes to source data toward easier informative integration into BIM models, by facilitating and enhancing interpretation needs. In this direction, Artificial Intelligence (AI) algorithms are part of the process. The adapted informative implementation of parametric models is based on digital source data (laser and photogrammetry) segmentation by AI according to specific topics (documentation, analysis, monitoring, conservation, project) and criteria (materials, techniques, components, structures).
AI-Enhanced UAV Photogrammetry Point Clouds for Cultural Heritage Assessment
Advancements in low-altitude remote sensing and image analysis have revolutionized the digitisation of real-world objects, initially represented as point clouds. Over the past decade, drone-based surveying has gained traction; however, noise during data capture and 3D reconstruction remains a critical challenge, affecting the accuracy and usability of UAV images in real-world applications. This study presents a novel approach to enhance cultural heritage assessment using deep learning models based on the Autoencoder Gaussian Mixture Clustering Model and the Autoencoder K-means Clustering Model. These models improve the accuracy of point clouds generated from UAV images for better condition assessment and survey. The research focuses on photogrammetric accuracy parameters for clustering point clouds. The new approach chooses K-means for finding global patterns due to its robust accuracy, and the Gaussian Mixture Clustering Model for local changes and inspection applications. Moreover, these models investigate the accuracy of point cloud clusters generated from K-means and the Gaussian Mixture Clustering Model. To validate the new approach, the study evaluates variations in key parameters under diverse built-environment conditions using the Temple of Neptune point cloud in Paestum, Italy. The results demonstrate significant improvements in point cloud 3D reconstruction, leading to more accurate surveys and assessments. The proposed method surpasses traditional techniques, offering enhanced applicability for point-cloud filtering. Furthermore, comparisons of clustering results highlight the algorithm, establishing its promising potential for advancing UAV-based surveying and inspection practices.
Semi-Automated Feature Recognition and Localization in 3D Point Cloud by Using AI
The application of Artificial Intelligence (AI) in Cultural Heritage surveying has gained notable interest, particularly in enhancing the classification and segmentation of 3D point clouds. Recent research focuses on automating processes to manage surveying data and integrating it into the BIM environment. The present research proposes a semi-automated AI-based pipeline to semantically classify architectural features in 3D point clouds and match them with an existing BIM library. The case study examines the school and theatre of the Crespi d’Adda industrial village, a UNESCO Cultural Heritage site. Data includes a 3D point cloud generated via terrestrial photogrammetry, with windows selected as the feature to model in the BIM library. The methodology encompasses three phases: (i) creating a Crespi d’Adda window dataset aligning with BIM library parameters; (ii) applying Machine Learning classifiers for semantic categorization; and (iii) using prediction algorithms to recognize windows in the point cloud, matching them to the BIM library, and calculating accuracy. This research bridges architectural representation and data mining, streamlining BIM reconstruction. Anticipated results include automated detection and labeling of elements with accurate placement in the BIM environment, enhancing efficiency and interdisciplinary integration.
From Temple to Church: The Evolution of San Lorenzo in Miranda Through Machine Learning
The paper addresses the challenges related to the application of machine learning solutions to support historical and architectural critical interpretation. The case study reported here is the complex of San Lorenzo in Miranda, located in the Roman Forum along the Via Sacra. The structure, originally conceived as a temple and later transformed into a church, is a multi-layered architectural palimpsest in which each construction phase, at least since Roman times, has inevitably influenced the subsequent modifications. The building’s rebirth upon itself through the integration and modification of its older portions is its main characteristic, imparting a specific complexity and interest. The first phase of the research focused on bibliographic study and 3D digital survey, both of which contributed to identifying the multiple construction phases of the building. In particular, the digital survey was implemented through two survey campaigns. The first involved a massive 3D laser scanner acquisition and UAV photogrammetric capturing, while the second integrated a topographic survey to georeference the captured data. The second phase concerned data interpretation, focusing on research questions related to a hypothetical reconstruction of the cella’s original wall covering layer. This question was addressed by leveraging machine learning algorithms used to automatically identify the covering traces.
Automated Recognition and Segmentation of Bricks Wall in Sicilian Monasteries
The oriental architecture present in the north-eastern cusp of Sicily is an expression of a mixed-race culture. This hybridization is evident in religious architecture. These buildings have the typological imprint of Catholic churches with Latin cross plans and towered facades, but preserve Arabic traces in the structure of the domes and connections, while also exhibiting Eastern Byzantine traditions in their masonry and rich decorations. The objective of this research is to automate the process of recognizing and segmenting bricks in wall structures to support the analysis of wall fabrics, a crucial task in archaeology and architectural restoration. Our approach processes a point cloud extracted from a facade to identify the wall texture. The results of segmentation can provide statistical information, documenting average brick size, mortar thickness, identification of homogeneous areas, and recognition of masonry sections built with different bricks. Alongside the numerical and abstract information, it is possible to identify the standard morphology of the brick, which also constitutes a sort of digital fingerprint of the church. The size of the bricks influences the geometry and layout of religious architecture. For example, the lateral facades are decorated with intertwined arches entirely composed of bricks. The spacing of the arches, their regularity, and the overall morphology are determined by the size and arrangement of the bricks. Bricks placed on the beds in different ways (stretcher, rowlock stretchers, or heading) determine the repetition or alteration of the geometric-formal modules reiterated in the elevations.
The Influence of Data Quality in Supervised ML-AI Classification Approaches for Historical Heritage
In recent years, the automatic segmentation and classification of digital survey data has been experimented with in built heritage studies. Despite the encouraging progress in the use of Machine and Deep Learning techniques, the semantic segmentation of point clouds is more complex, especially for the historic environment for which, due to the heterogeneity of shapes, it is more difficult to recognize homogeneous regions with similar properties. Given the need to process a large volume of already annotated data for the training and recognition of new scenes, the type and quality of the initial data play a fundamental role in the classification process, as they influence the subdivision into predefined categories that are not always consistent with a decomposition into architectural elements and sub-elements shared by the scientific community. This is an interpretative problem that already emerges from traditional manual labelling, which, being highly subjective, reduces the reproducibility of the results. The paper focuses on understanding to what extent the recognition of homogeneous regions is influenced by factors such as: manual labelling carried out by annotators with different specializations; density value of the point clouds; and type of data acquired depending on the acquisition sensor. These evaluations were conducted by employing the Random Forest algorithm on specific pre-processed point cloud datasets, with reference to the typology of the Franciscan cloister, in order to make the recognition flows more controlled and less ambiguous, aiming at an advancement towards more efficient modelling and management of the existing architectural heritage.
Impact of Varying Street View Perspectives on Urban Perception: The Case of Celoria Street in Milan
Urban environments significantly influence people’s perception and walkability. Advances in computer vision and the availability of open-source Street View Imagery (SVI) have increased the use of Google Street View (GSV) for perceptual predictions and walkability assessments. However, a critical issue arises from the discrepancies between GSV images, captured from street centerlines, and SVI taken from pedestrian perspectives on sidewalks. This study examines whether people’s perceptions and street element proportions derived from GSV images align with those from sidewalk viewpoints, providing a more accurate basis for urban studies. Taking Celoria Street in Milan as a case study, two sets of 360° panoramic images were collected, one from the street center and the other from the sidewalks. These images were processed using a pre-trained perception prediction model and image segmentation techniques to generate perception responses. Dynamic Time Warping (DTW) was applied to assess the consistency between the two datasets, while Ordinary Least Squares (OLS) regression was used to analyze the impact of viewpoint changes along the street scene. Findings indicate that differences in sampling perspectives can affect urban environment assessment and perception predictions. This study highlights the potential biases of GSV data for analyzing urban environments and perceptions, advocating for more cautious use of SVI to ensure robust predictions on urban perception and walkability.
The New A.I.: Gaining Control Over the Noise
The paper investigates the evolution of AI-assisted architectural representation from early unpredictable text-to-image generation systems toward more controlled and precise workflows based on Stable Diffusion, ControlNet, and LookX AI. The research analyzes the role of latent spaces, diffusion models, GANs, convolutional neural networks, and AI rendering systems in architectural visualization, emphasizing the transition from exploratory AI image production to controllable design-oriented generation. Through experimental workflows combining Rhinoceros 3D models, ControlNet preprocessors, segmentation maps, depth maps, edge detection, and prompt engineering, the study evaluates how AI systems can support architectural rendering, stylistic control, spatial coherence, and atmosphere generation. The paper compares open-source and cloud-based AI platforms, discussing the balance between creativity, predictability, customization, and architectural precision in contemporary AI-assisted design workflows.
Comparative Analyses Between Sensors and Digital Data for the Characterization of Historical Surfaces
The paper investigates the use of radiometric and colorimetric data derived from terrestrial laser scanning for the characterization and interpretation of historical architectural surfaces. The research focuses on the relationship between intensity values, materials, construction techniques, and surface decay pathologies, exploring how reflectance data can support semantic segmentation and machine learning classification processes in cultural heritage documentation. Through comparative analyses between different laser scanner sensors and manual segmentation of point clouds, the study evaluates the reliability of intensity values as discriminative features for material recognition and conservation assessment. The workflow integrates point cloud processing, histogram analysis, comparative diagrams, and surface feature interpretation to support future AI-assisted classification and Scan-to-BIM applications for heritage conservation.
A Proposal of Integration of Point Cloud Semantization and VPL for Architectural Heritage Parametric Modeling
The paper proposes a reproducible workflow for the semi-automated parametric modeling of architectural heritage starting from point cloud data acquired through laser scanning and photogrammetry. The methodology integrates automatic semantic segmentation, geometric feature recognition, mathematical analysis in Matlab, and Visual Programming Language (VPL) procedures in Dynamo to support Scan-to-BIM workflows. The system identifies geometric primitives from segmented point clouds, extracts modeling parameters, and generates parametric vault models automatically through a VPL environment. The research aims to reduce subjectivity, modeling time, and manual effort in heritage HBIM processes while improving reproducibility and scalability for architectural documentation and conservation.