| Title |
Comparative Study on the Performance of Multi-view Image Learning-based Penetrability Analysis Model for Classifying Irrelevant Clashes in BIM model |
| Authors |
Hyunwoo Lee ; Youngsu Yu ; Wonbok Lee ; Bonsang Koo |
| DOI |
https://dx.doi.org/10.6106/KJCEM.2025.26.5.026 |
| Keywords |
Penetrability; Sleeve installation; Multi-view Vision Transformer; Multi-View CNN; BIM |
| Abstract |
BIM models are developed in a fragmented manner by discipline and integrated during the detailed design phase, leading to numerous clashes. Many of these are irrelevant clashes that do not require intervention, yet significant effort is needed for classification. Sleeve installation feasibility assessment is a key task in this process, and various studies have explored automated penetrability analysis. However, existing methods relied on manually defined inference rules or failed to capture detailed clash patterns. This study developed penetrability analysis models using Multi-view CNN (MVCNN) and Multi-view Vision Transformer (MVT), both of which enable multi-view image training. Experimental results showed that MVT achieved an accuracy (ACC) of 0.98, outperforming MVCNN by 0.13 ACC. MVT’s superiority was attributed to its attention mechanism, which focused on clash-prone regions, unlike MVCNN’s emphasis on overall object geometry. These findings demonstrate practical value by enabling the early identification of irrelevant clashes to enhance design efficiency and accuracy, while supporting more precise construction planning through detailed penetration information. |