Daejeon, July 27–31
A current challenge of using 3D models in museums, education or tourism is limited availability of cultural heritage 3D models (Münster et al. 2025). The Digital Europe project 3DBigDataSpace addresses this gap. A consortium of ten partners — coordinated by Time Machine Organisation together with the Friedrich-Schiller-Universität Jena, the Poznań Supercomputing and Networking Center (PCSS), the University of Vigo (UVigo), the Bruno Kessler Foundation (FBK-3DOM), the SMEs IN2 and Rooom, and the Hunt Museum Limerick — assesses, retrieves, and openly releases 3D heritage assets for aggregation into the common European dataspace for cultural heritage. The work presented in this article is about first steps (1) to compile a large-scale pool of 3D assets of cultural heritage and ready-to-use viewer applications and (2) to enrich and employ them in various settings to enhance and validate the usability of 3D models. While the data collection and viewer development have already reached an intermediate state, the validation within use cases is currently ongoing. An extended version of this abstract can be found in (Münster et al. 2026)(Münster et al. 2025). The Digital Europe project 3DBigDataSpace addresses this gap. A consortium of ten partners — coordinated by Time Machine Organisation together with the Friedrich-Schiller-Universität Jena, the Poznań Supercomputing and Networking Center (PCSS), the University of Vigo (UVigo), the Bruno Kessler Foundation (FBK-3DOM), the SMEs IN2 and Rooom, and the Hunt Museum Limerick — assesses, retrieves, and openly releases 3D heritage assets for aggregation into the common European dataspace for cultural heritage. The work presented in this article is about first steps (1) to compile a large-scale pool of 3D assets of cultural heritage and ready-to-use viewer applications and (2) to enrich and employ them in various settings to enhance and validate the usability of 3D models. While the data collection and viewer development have already reached an intermediate state, the validation within use cases is currently ongoing. An extended version of this abstract can be found in (Münster et al. 2026)
For the 3D content, large-scale datasets such as Objaverse including 10,2 million 3D models (Deitke et al. 2023) or ShapeNet counting 50k 3D assets (Deitke et al. 2023) or ShapeNet counting 50k 3D assets (Chang et al. 2015), as well as repositories like Sketchfab hosting several 100.000 heritage items (Chang et al. 2015), as well as repositories like Sketchfab hosting several 100.000 heritage items (Flynn 2019) have been compiled and shared. At the same time, large-scale 2D/3D datasets such as MVImgNet2 (Flynn 2019) have been compiled and shared. At the same time, large-scale 2D/3D datasets such as MVImgNet2 (Wu et al. 2024) and MegaScenes (Wu et al. 2024) and MegaScenes (Tung et al. 2024) were created. In addition, campaigns as Scan the World or Global Digital Heritage propose a collective effort to digitize cultural heritage at large scale. Various research investigated the retrieval of sufficient metadata from 3D objects and metadata to spatialize and temporalize this material (e.g. (Tung et al. 2024) were created. In addition, campaigns as Scan the World or Global Digital Heritage propose a collective effort to digitize cultural heritage at large scale. Various research investigated the retrieval of sufficient metadata from 3D objects and metadata to spatialize and temporalize this material (e.g. (Orzechowski et al. 2025)). (Orzechowski et al. 2025)).
Between 11/2023 and 04/2025 we harvested metadata via Python and PHP server-side scripts feeding an SQL/Unix file store, gathering 11.4 M records (Münster 2026). Selection was applied in three explicit, documented layers:(Münster 2026). Selection was applied in three explicit, documented layers:
The resulting study subset of ca. 130.000 models is listed in Table 1.
Table . Processed datasets.
The Heritage Data Processor (HDP) (Ukolov 2025) is a GUI / API / CLI application for the processing, enrichment, and management of digital heritage assets. Multimodality (mesh + photograph + text) is a primary design constraint. Its modular pipeline architecture allows project partners to plug in services as independent components. Each module declares its inputs, outputs, and runtime dependencies; pipelines are composed declaratively and executed under Docker. The framework combines machine -learning models, large language models (LLMs), vision-language models (VLMs), and external APIs, and uses Zenodo as a stable storage endpoint, ensuring FAIR-aligned long-term preservation alongside open sharing. Using the HDP and its predecessor Zenodo Toolbox, approximately 65.000 3D mesh models and additional 65.000 photographs have been processed and stored in Zenodo so far.(Ukolov 2025) is a GUI / API / CLI application for the processing, enrichment, and management of digital heritage assets. Multimodality (mesh + photograph + text) is a primary design constraint. Its modular pipeline architecture allows project partners to plug in services as independent components. Each module declares its inputs, outputs, and runtime dependencies; pipelines are composed declaratively and executed under Docker. The framework combines machine -learning models, large language models (LLMs), vision-language models (VLMs), and external APIs, and uses Zenodo as a stable storage endpoint, ensuring FAIR-aligned long-term preservation alongside open sharing. Using the HDP and its predecessor Zenodo Toolbox, approximately 65.000 3D mesh models and additional 65.000 photographs have been processed and stored in Zenodo so far.
A MeshOptimizer tool has been developed within the project activities to facilitate large-scale 3D mesh processing and improve models’ visualization across various platforms. The tool is a fully automated and modular 3D mesh processing pipeline, designed for optimizing both geometry and visual texture quality. The framework supports both decimation and remeshing, as well as texture improvement through the generation of PBR (Physically Based Rendering) maps, including diffuse, ambient occlusion, roughness, and normal maps. Geometry optimization is designed to balance mesh simplification and object fidelity. The framework, containerized via Docker and implemented in Python and C++, leverages Blender for texture processing and geometry decimation, xatlas for efficient UV mapping, and C++ libraries (such as PMP-library and CGAL) for remeshing, surface analysis, and geometric transformations. The Hausdorff distance is iteratively computed to analyze the geometric deviation at each remeshing stage and ensures adaptive optimization while preserving the integrity of the original model (see Figure 1).
Fig. . An example of mesh simplification obtained with the MeshOptimizer tool.
Enrichment proceeds in three stages: (a) text pre-processing (language detection via Fasttext; translation via NLLB-distilled-1.3B; short-description summarisation via distilbart-cnn-12-6; image captioning via LLAVAnext / Mistral 7B); (b) categorisation against Europeana CHO classes; (c) geolocalization, and (d) linking to other open data providers as Wikidata and GettyAAT. For each stage (Table 2) we benchmarked alternative models — e.g. against manual labelled data or a Europeana ground-truth subset of 8.708 validated items as records carrying expert-curated CHO type and country/place metadata. Inputs are titles, translated descriptions, and rendered images of the mesh; outputs are predicted CHO type, country, region, and city, with confidence scores.
For content classification (categorizing and labelling data) three complementary approaches were developed. (1) An image-based MMLM was used for categorization with results manually verified to provide ground-truth data. (2) A multimodal classifier combining results from different LLM and VLM for cross validation was deployed for unattended classification. (3) For the labelling of parts of models a point cloud-based ML using a random forest classification did take place. All models were easy to be integrated as Docker containers.
Table . Models tested for metadata enrichment.
For geolocalization (1) a manual geolocalisation tool enabled accurate positioning of 3D models on 3D terrain map representations. Geographic features are saved to an IIIF Presentation v4 Manifest with seamless Zenodo integration (2) a LLM-/VLM-based approach combined rough localisation followed by satellite -imagery-based fine pose estimation (Rigon et al. 2026). (3) An alternative approach was particularly used to retrieve geolocalizations from textual captions, following a multi staged process. Via this approach 186.896 datasets were processed (Figure 2). (Rigon et al. 2026). (3) An alternative approach was particularly used to retrieve geolocalizations from textual captions, following a multi staged process. Via this approach 186.896 datasets were processed (Figure 2).
Fig. 2. Geocoded 3D models and point clouds per data source in Europe (n = 78.901).
With regards to the performance of specific LLMs Qwen 3.0 (14B) achieved the strongest overall results, outperforming even larger models. The geographic distribution of datasets reveals a shift from earlier Mediterranean dominance to broader European coverage, with Germany, France, the Netherlands, and the UK now among the leading countries. This trend strongly correlates with the availability of openly licensed photographic content on Wikimedia Commons, while national legal frameworks—such as restrictions on cultural heritage imagery in Greece—directly affect dataset representation. (Münster 2026). (Münster 2026).
A set of currently thirteen viewers is integrated through the DFG 3D Viewer wrapper to view the models according to specific needs (Münster et al. in press). (Münster et al. in press).
Those viewers include ATON, 3D-HOP, Kompakkt, model-viewer, and the AIM Viewer. A set of viewers has been specifically developed within the 3DBigDataSpace project (Fig. 3). Those include the PCSS Viewer (React + Three.js, .glb/.gltf, annotations), the 4D Browser and 4D City World Viewers (time-variant 2.5D / 3D city visualisation on mobile and desktop), a Holopyramid viewer and an XR / AR Viewer (Rooom AG). As standalone applications the web-based ArtefactIQ application (IN2) integrates these viewers behind a CMS-like authoring interface so that cultural-heritage professionals, educators, and students can build small games and exhibits without writing code. The UVigo routing app https://camino-57345.web.app/, accessed on 6th May 2026.
Fig. 3. Viewers application interfaces.
The use of XR applications has a significant effect on museum learning (Xu et al. 2023), enjoyment (Xu et al. 2023), enjoyment (Dong et al. 2024) and engagement. The 3D models has been used in the Horizon Europe project META-MUSEUM which investigates how people experience cultural heritage by fostering empathy, co-creation, and inclusion. To validate those affective effects museum cooperations took part with four museums so far. The datapool is used to creating a set of prototypes. Specific scenarios include (1) to validate affective and learning effects of remote and physical representations, (2) provide content in a remote environment vs. on site, (3) test different ways of storytelling, e.g. with regards to multi-coding (Dong et al. 2024) and engagement. The 3D models has been used in the Horizon Europe project META-MUSEUM which investigates how people experience cultural heritage by fostering empathy, co-creation, and inclusion. To validate those affective effects museum cooperations took part with four museums so far. The datapool is used to creating a set of prototypes. Specific scenarios include (1) to validate affective and learning effects of remote and physical representations, (2) provide content in a remote environment vs. on site, (3) test different ways of storytelling, e.g. with regards to multi-coding (Paivio 2006), varying narratives and customization (Paivio 2006), varying narratives and customization (Shen et al. 2024). Several test settings have been compiled to enable lab tests of different modes of interaction with content. A first cohort of proband tests has been conducted since early 2026 and revealed that 3D models need to be accompanied with information about the shown object in order to be understood by visitors.(Shen et al. 2024). Several test settings have been compiled to enable lab tests of different modes of interaction with content. A first cohort of proband tests has been conducted since early 2026 and revealed that 3D models need to be accompanied with information about the shown object in order to be understood by visitors.
Educational courses comprise school courses designed by students of art history and history teaching, which were offered to primary and secondary school pupils in Jena in the context of school and extracurricular working groups (Münster et al. 2025). These courses did produce content for the 4D world viewers – e.g. 3D scans of city sculpture, textual descriptions of landmarks and virtual city tours for children. In total 180 pupils participated in these educational courses so far. The Hunt museum in Limerick used the ArtefactIQ tool to augment the existing learning material on (Münster et al. 2025). These courses did produce content for the 4D world viewers – e.g. 3D scans of city sculpture, textual descriptions of landmarks and virtual city tours for children. In total 180 pupils participated in these educational courses so far. The Hunt museum in Limerick used the ArtefactIQ tool to augment the existing learning material on Life in the Bronze Age for secondary students starting in November 2025. Within a gamified experience where students aim to recognise certain characteristics of museum artefacts (for e.g. to assess the period of their origin), 3D models and their annotations play a key role. https://www.huntmuseum.com/schools-resources/pre-christian-ireland-3d-big-data, accessed on 6th May 2026.
Virtual Travelguides are well-suited to provide tourists with information about their route and surroundings. The UVigo routing app is designed to digitally enhance the cultural and touristic experience for pilgrims on their pilgrimage routes (i.e., Camino Miñoto Ribeiro, Camino de Santiago, and Roma) through immersive technologies. Built on the model_viewer_plus library together with consortium-supplied 3D assets and native AR engines (ARCore on Android, ARKit on iOS) via ar_flutter_plugin_2, the app lets users navigate heritage sites and landmarks in a fully interactive 3D environment and overlay cultural elements in real space for an immersive pilgrimage-route experience.
Functionalities include AI-assisted POI enrichment — using Qwen3-8B for multilingual text and translation, FLUX.1-schnell for image generation when none is supplied, and automated moderation to filter spam or inaccurate content — alongside optimized 3D-model handling through pre-caching, dynamic loading, and the integration of additional assets from cultural consortia. In parallel, the project does explore an integration with the Infraestructura de Datos Espaciales de España (IDEE) to enrich POI metadata with authoritative geospatial information; this remains exploratory given current accessibility and interoperability constraints, but a successful link-up would enable more accurate, dynamic, and standardized geospatial referencing of cultural heritage routes within the national infrastructure ecosystem.
Large-scale, openly licensed 3D cultural heritage content can be aggregated and enriched on the scale of the European Data Space, but only if the selection, processing and viewing infrastructures are explicitly modular, multilingual and bias-aware. This includes identifying a predominance of built heritage and artworks, such as sculptures, which are mainly located in Central Europe and the UK. Regarding technologies, text-based enrichment currently outperforms image-based methods for geolocation and classification, and mid-size LLMs such as QWEN 14B outperform larger and smaller LLMs, as well as transformer baselines, for this task. Application testing of all scenarios is ongoing in spring 2026. Preliminary results revealed the need to enrich 3D views with information, as well as showing that interactive 3D enhances the duration and intensifies the learning effect for learners. The next steps include pipeline orchestration and consolidation of the technical components. Until the project's completion in July 2026, the consortium will continue to validate the workflow in real-world settings.
The research presented here was carried out in the EU-funded projects 3DBigDataSpace (Grant No. 101173385) and MetaMuseum (Grant No. 101132488). We thank all consortium partners and the participating museums.
CHANG, A. X., FUNKHOUSER, T., GUIBAS, L., HANRAHAN, P., HUANG, Q., LI, Z., SAVARESE, S., SAVVA, M., SONG, S. & SU, H. 2015. Shapenet: An information-rich 3d model repository. CHANG, A. X., FUNKHOUSER, T., GUIBAS, L., HANRAHAN, P., HUANG, Q., LI, Z., SAVARESE, S., SAVVA, M., SONG, S. & SU, H. 2015. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 2, 3, 9.
DEITKE, M., LIU, R., WALLINGFORD, M., NGO, H., MICHEL, O., KUSUPATI, A., FAN, A., LAFORTE, C., VOLETI, V., GADRE, S. Y., VANDERBILT, E., KEMBHAVI, A., VONDRICK, C., GKIOXARI, G., EHSANI, K., SCHMIDT, L. & FARHADI, A. 2023. Objaverse-XL: A Universe of 10M+ 3D Objects. arXiv, 2307.05663.
DONG, H., CAI, Y. & WEI, Q. 2024. The Impact of VR Museum Exhibition Design on User Emotional Experience: A Meta-Analysis. Proceedings of the 2024 International Conference on Cloud Computing and Big Data. Dali, China: Association for Computing Machinery.
FLYNN, T. 2019. Over 100,000 Cultural Heritage Models on Sketchfab [Online]. Available: https://sketchfab.com/nebulousflynn/collections/over-100000-cultural-heritage-models-on-sketchfab [Accessed 30.1.2026].
MüNSTER, D. L., MüNSTER, S. & DIETZ, R. 2025a. Heritage education projects by university students for pupils: Digital and data-driven learning in a humanities teaching-learning lab. Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLVIII-M-9-2025, 1037–1042.
MüNSTER, S. 2026. Large-Scale Metadata Processing for 3D Cultural Heritage Objects. Land, 15, 751.
MüNSTER, S., KUROCZYNSKI, P., MOSCH, M., BAJENA, I., BECK, C., WEIGELT, M. & DWORAK, D. in press. Implementing a distributed infrastructural workflow for 3D Models in Germany. Achievements of the DFG 3D-Viewer Project's Second Funding Phase. In: MüNSTER, S., KUROCZYńSKI, P. & APOLLONIO, F. (eds.) Research and Education in Urban History in the Age of Digital Libraries 2025. Cham: Springer.
MüNSTER, S., MARCINIAK, P., WęGRZYNOWSKI, M., DE ARRIBA PéREZ, F., GARCíA MéNDEZ, S., UKOLOV, D., RIGON, S., FARELLA, E. M., REMONDINO, F., SUN, Y. & STAN, A. 2026. Workflows for analysing and utilizing large-scale 3D datasets of cultural heritage. Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLVIII-2/W12-2026, 319–325.
MüNSTER, S., MEDICI, M., FRESA, A. & STAN, A. 3D Data in the Data Space for Cultural Heritage. In: CAMPANA, S., FERDANI, D., GRAF, H., GUIDI, G., HEGARTY, Z., PESCARIN, S. & REMONDINO, F., eds. Digital Heritage 2025, 2025b Sienna.
ORZECHOWSKI, M., OPIOłA, Ł., MARTíNEZ, I. L., IOANNIDES, M., PANAYIOTOU, P. N., DUTKA, Ł., SłOTA, R. G. & KITOWSKI, J. 2025. Integrated data, metadata, and paradata management system for 3D Digital Cultural Heritage objects: Workflow automation, federated authentication, and publication. Future Generation Computer Systems, 107964.
PAIVIO, A. Dual Coding Theory and Education (Draft). Pathways to Literacy Achievement for High Poverty Children,” The University of Michigan School of Education, 29. 9.- 1. 10. 2006, 2006.
RIGON, S., FARELLA, E. M., MORELLI, L., BERTOLASI, G., REMONDINO, F. & MüNSTER, S. 2026. 3DGeoRef: an automated framework for georeferencing heritage 3D models. Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLVIII-2/W12-2026, 407–414.
SHEN, J., MIRE, J., PARK, H. W., BREAZEAL, C. & SAP, M. 2024. HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs.
TUNG, J., CHOU, G., CAI, R., YANG, G., ZHANG, K., WETZSTEIN, G., HARIHARAN, B. & SNAVELY, N. 2024. MegaScenes: Scene-Level View Synthesis at Scale. https://arxiv.org/abs/2406.11819, 2406.11819.
UKOLOV, D. 2025. Heritage Data Processor (HDP): A Modular Architecture for Processing and Persistently Storing Multimodal Cultural Heritage Data. Working Paper [v1]. Zenodo, 10.5281/zenodo.17643051.
WU, Y., SHI, L., LIU, H., LIAO, H., QIU, L., YUAN, W., GU, X., DONG, Z., CUI, S. & HAN, X. 2024. MVImgNet2.0: A Larger-scale Dataset of Multi-view Images. ACM Trans. Graph., 43, Article 173.
XU, W., DAI, T.-T., SHEN, Z.-Y. & YAO, Y.-J. 2023. Effects of technology application on museum learning: A meta-analysis of 42 studies published between 2011 and 2021. Interactive learning environments, 31, 4589–4604.