ArchShapesNet dataset

 A collection of 13 types of BIM elements with high utility in the architectural domain

• 4,000 elements per type, totaling 52,000


• 12 images per element


• Dataset used for object classification (using MVCNN)

 ArchShapesNet overview  


The ArchShapesNet is an architectural BIM element dataset developed by i3 lab at Seoul National University of Science and Technology. It was developed for training deep learning algorithms, especially Multi-view CNN (MVCNN). The dataset consists of 4,000 BIM elements per 13 commonly used types, totaling 52,000 elements. These elements have been collected from four IFC standard building models and several online BIM repositories (KBIMS Library, NBS, bimobject, etc.). It also includes elements that have been augmented through parametric adjustments using Revit Dynamo. The dataset is composed of 12 images for each individual element, which is required to train MVCNN.

  ArchShapesNet is free and open to download for research purposes, and we hope the dataset contributes to the advancement of BIM and AI in the AEC industry.

BIM model

BIM library




KBIMS Library


Office building

Single house





Cultural and assembly facilitiesEducational research facilitiesSTANLEY

 Starting point

  A critical aspect of BIM is the capability to embody semantic information about its element constituents. To be interoperable, such information needs to conform to the Industry Foundation Classes (IFC) standards and protocols. Working under the IFC protocol requires BIM elements, relationships, and their properties to be represented in conformance to its standards. However, due to the lack of logical rigidity of the IFC schema, IFC model instances are prone to misrepresentations and misinterpretations, resulting in a lack of semantic integrity.

  Artificial intelligence approaches have been explored as a way to verify the semantic integrity of BIM to IFC mappings by learning the geometric features of individual BIM elements. Specifically, the i3 lab research team explored the use of different machine learning and deep learning models to determine their applicability (Koo et al., 2019;  Koo et al., 2021; Koo et al., 2021). Mainly, we trained learning models based on the geometric features of individual BIM elements. We attained the most promising results from incorporating Multi-View CNN (MVCNN), a geometric deep learning model that learns from multiple panoramic images of a 3D artifact to learn and distinguish its shape (Su et al., 2015).


  [Semantic Enrichment Research in i3LAB]

 Dataset development

  However, despite its relative high performance, MVCNN still was limited in correctly classifying specific BIM elements despite their geometric differences. The classification errors were attributed mainly to the limited number of training data, as well as the imbalance in the number of samples per element class. 
  Initially, BIM element samples were collected from open-source libraries or existing BIM models. However, such avenues were limited due to copyright issues or lack of usable samples from the attained BIM models. 

  Thus, a data augmentation process for creating additional samples was developed employing parametric modeling. Specifically, we created new samples by modifying the parameters of the original BIM elements. By randomly ‘tweaking’ these parameters within their permitted ranges, we were able to make legitimate duplicates. The table below shows the parameters modified and their corresponding boundary ranges. The augmentation process was implemented using Revit Dynamo.

[Parameters and range for each BIM element]

[Parametric modeling using Revit Dynamo]



 BIM element types and statistics

The ArchShapesNet consists of 13 element types that were found to be commonly used in the architectural field. Sample images and distribution of numbers by element are shown in the table and figure below.

[BIM elements Image of ArchShapesNet]


           [Element breakdown of ArchShapesNet]


The basic dataset (Arch-3 version) includes three types of BIM elements: doors (single, double, and revolving), walls, and windows. Please click on the ‘Arch-3 version’ button below and fill in the Google form. We will send you the download link in 7 business days once you have submitted the form. If you can't open the form, please send an email to i3 lab (email: to make a  direct request.

  Please note that
ArchShapesNet is currently available for research purposes only, and no commercial use is permitted. 

  If you need the full dataset, please click on the
‘Arch-13 version’ button below to make a download request. These requests will be reviewed on a case-by-case basis. We suggest you try the Basic Dataset in your research/projects before requesting the full dataset.



•   3,000 BIM elements

•   3 types of BIM elements; doors, walls, windows

•   1,000 elements per type



 •   52,000 BIM elements

 •   13 types of BIM elements

 •   4,000 elements per type


ArchShapesNet leaderboard

 The classification performance of the deep learning models trained and tested on ArchShapesNet is shown below. Please feel free to send your classification results, together with the metrics, and we will record them on the leaderboard below.

  In addition, we ask that if you use
ArchShapesNet in your research, you specify the source in your reference as follows: Yu, Y. S., Lee, K. E., Ha, D. M., Koo, B. S. (2021). Enhancing Deep Learning-based BIM Element Classification via Data Augmentation and Semantic Segmentation, 2021 Proceedings of the 38th ISARC, pp. 227-234.


 Accuracy (%)














Contribute to ArchShapesNet

ArchShapesNet needs your support to develop a more comprehensive dataset. Please contact i3 lab (email: if you wish to contribute additional BIM element samples or BIM models (in IFC format) to ArchShapesNet. If you wish to conduct research collaboratively using ArchShapesNet, please send an email directly to  Professor Bonsang Koo (


  Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.