This thesis investigates the limits and potential of artificial intelligence in the geometric reading of architectural space from monocular images. The research starts from descriptive geometry and NURBS-based modelling in order to define a critical framework for comparing geometric abstraction and statistical abstraction. The experimental phase uses controlled models in which coordinates, planes, lines, intersections, camera parameters and ground truth are known in advance. The results show that multimodal systems often produce plausible descriptions, but not reliable metric reconstructions. The applied chapter transfers these findings to a ComfyUI workflow, understood not as an automatic surveying tool, but as a mediation environment: it segments, isolates, reprocesses and generates intermediate representations that can be documented and assessed. The thesis concludes that AI can support architectural representation only when it is embedded in a guided, verifiable procedure based on explicit geometric constraints.
La tesi indaga i limiti e le possibilità dell’intelligenza artificiale nella lettura geometrica dello spazio architettonico a partire da immagini monoculari. Il lavoro muove dalla tradizione della geometria descrittiva e dalla modellazione NURBS per costruire un criterio di confronto tra astrazione geometrica e astrazione statistica. La sperimentazione procede attraverso modelli controllati, nei quali coordinate, piani, rette, intersezioni, camera e ground truth sono definiti a priori. I risultati mostrano che i sistemi multimodali producono descrizioni spesso plausibili, ma non ricostruzioni metriche affidabili. Il capitolo applicativo trasferisce tali conclusioni a un workflow ComfyUI, inteso non come strumento di rilievo automatico, ma come ambiente di mediazione: segmenta, isola, rielabora e genera rappresentazioni intermedie documentabili. La ricerca conclude che l’IA può contribuire alla rappresentazione architettonica solo se inserita in una procedura guidata, verificabile e fondata su vincoli geometrici espliciti.
Ricostruzioni architettoniche 3D da immagini monoculari tramite intelligenza artificiale : sperimentazione e analisi di workflow generativi in ambiente ComfyUI
Aria, Arvin
2025/2026
Abstract
This thesis investigates the limits and potential of artificial intelligence in the geometric reading of architectural space from monocular images. The research starts from descriptive geometry and NURBS-based modelling in order to define a critical framework for comparing geometric abstraction and statistical abstraction. The experimental phase uses controlled models in which coordinates, planes, lines, intersections, camera parameters and ground truth are known in advance. The results show that multimodal systems often produce plausible descriptions, but not reliable metric reconstructions. The applied chapter transfers these findings to a ComfyUI workflow, understood not as an automatic surveying tool, but as a mediation environment: it segments, isolates, reprocesses and generates intermediate representations that can be documented and assessed. The thesis concludes that AI can support architectural representation only when it is embedded in a guided, verifiable procedure based on explicit geometric constraints.| File | Dimensione | Formato | |
|---|---|---|---|
|
Arvin Aria.pdf
non accessibile
Descrizione: Tesi in formato PDF
Dimensione
7.89 MB
Formato
Adobe PDF
|
7.89 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/260535