Large-Scale Georeferenced 3D Reconstruction of Pipe Organs from Multi-Source Heritage Data: An End-to-End Pipeline

Large-Scale Georeferenced 3D Reconstruction of Pipe Organs from Multi-Source Heritage Data: An End-to-End Pipeline Working Paper [v1] 2025 Dominik Ukolov Digital Humanities (Image/Object), Friedrich-Schiller-University Jena Research Center DIGITAL ORGANOLOGY, Leipzig University [email protected] Working Paper [v1] submitted on 17 November, 2025. DOI [v1]: 10.5281/zenodo.17629438 Abstract Cultural heritage documentation faces critical challenges regarding the integration of fragmented multimodal data sources, specifically images, textual metadata, and spatial information, into unified and actionable digital representations. Existing approaches frequently treat 3D reconstruction, georeferencing, and data management as isolated tasks, thereby limiting scalability and downstream applications in heritage preservation and research. An end-to-end pipeline is presented here for the large-scale georeferenced 3D reconstruction of pipe organs and church interiors from heterogeneous heritage data, designed for direct applications in acoustical modeling and broader cultural heritage domains. The proposed framework integrates (1) multi-source image retrieval and preprocessing; (2) neural 3D reconstruction via Gaussian Splatting and related generative methods; (3) LLM-powered semantic georeferencing utilizing textual metadata and external APIs; (4) OSM-based metric scaling and alignment; and (5) the extension of the MODAVIS Data Model. This data model serves as a multimodal database architecture linking organs, persons, locations, multimedia assets, and controlled vocabularies within a unified knowledge graph. By automating the transformation of dispersed archival materials into metrically accurate, georeferenced 3D models, the system is designed to provide geometric foundations for room acoustics simulation while enabling semantic queries across heritage collections. This working paper details the architectural design and integration of these components, establishing the mechanism for LLM-based toponymic disambiguation, the strategy for Gaussian Splatting of complex heritage objects, and the protocol for OSM-constrained scaling. Proof-of-concept implementations on selected pipe organ case studies are described to illustrate the feasibility of the pipeline stages. This integrated approach bridges computer vision, natural language processing, and heritage informatics, offering a scalable methodology for digital heritage preservation and interdisciplinary research applications. Keywords 3D Reconstruction ·Digitization ·Virtualization ·Musical Instrument ·Pipe Organ Large-Scale 3D Reconstruction of Pipe Organs Working Paper [v1] 1 Introduction Recent advances in the field of 3D reconstruction, particularly those based on Gaussian Splatting, have rendered the large-scale processing of physical entities feasible within established resource constraints. However, in order to avoid generating indiscriminate data that is difficult to find and classify, correct classification, annotation, and accessible storage are essential. Besides the inherent challenges, the object class of pipe organs exhibits numerous advantages for this task, including its location-specific nature and historical persistence, its presence in popular tourist destinations, its comprehensive documentation, and its status as a paradigmatic example of spatiotemporal object state changes. The MODAVIS dissertation project is developing a data model specifically for pipe organs that is capable of mapping all information and intermodal links like images, making it ideally suited to meet the demands of large-scale 3D reconstruction. The following paper delineates the fundamental components of the reconstruction pipeline and aligns these with the data model. It provides a detailed illustration of the processing of data provenance and exemplifies the modularity of reconstruction algorithms, as this field is rapidly evolving. A test data set consisting of five pipe organ objects was created and used to demonstrate the application of the concept and its scalability. In the subsequent research endeavours within the scope of the dissertation project, it is envisaged to augment the quantity and heterogeneity of the target objects by a substantial margin, with the objective of attaining a minimum of 1,000. 2 Related Work 2.1 Image-Based 3D Reconstruction Image-based 3D reconstruction has traditionally been framed as a two-stage pipeline comprising structurefrom-motion (SfM) for joint camera pose and sparse point cloud estimation, followed by multi-view stereo (MVS) for dense surface reconstruction [1]. This paradigm is embodied in mature software frameworks such as COLMAP, which support unordered image collections, robust feature matching, and scalable bundle adjustment, and therefore remain a de facto reference for evaluating newer learning-based approaches [1]. Bridging the gap between classical feature engineering and end-to-end reconstruction, deep learning-based feature extractors and matchers have significantly improved robustness in challenging environments. Methods such as SuperPoint and SuperGlue leverage graph neural networks to establish correspondences across wide baselines and substantial illumination changes, scenarios where traditional descriptors like SIFT often fail [2]. These learned representations are particularly valuable for heritage datasets, where historical imagery must be registered against modern ground truth despite degradation or stylistic differences. Recent work has increasingly combined classical geometric optimisation with deep neural representations, ranging from learned feature descriptors and matchers to implicit scene models such as neural radiance fields (NeRFs) and their real-time variants [3]. Gaussian Splatting, in particular, enables high-quality rendering of complex scenes from dense image sets at interactive frame rates, and has been shown to handle challenging materials and lighting conditions that arise frequently in cultural heritage documentation [3]. Comprehensive surveys of the field trace the evolution from handcrafted descriptors to deep learning, noting that while detector-free matchers like LoFTR effectively handle low-texture regions, they introduce significant computational overhead [4, 5]. To mitigate this, recent pipelines increasingly adopt hybrid approaches or lightweight transformer variants to balance the dense matching capability of learning-based methods with the efficiency required for large-scale datasets [4]. More recently, fully feed-forward architectures have emerged that directly predict dense depth, point maps, and camera parameters from monocular or multi-view inputs, thereby reducing reliance on iterative geometric optimisation. Visual Geometry Grounded Transformers (VGGT) infer camera poses and dense geometry in a single forward pass and achieve state-of-the-art performance on several 3D benchmarks, which makes them an attractive baseline for large-scale, heterogeneous reconstructions [6]. Similarly, permutation-equivariant models such as π3 remove the need to designate a fixed reference view, 1 Large-Scale 3D Reconstruction of Pipe Organs Working Paper [v1] improving robustness to view ordering and enabling efficient inference on large image sets [7]. In parallel, geometry-centric methods like DUSt3R learn to reconstruct accurate 3D point maps and camera poses from sparse or unstructured imagery, and have already demonstrated strong performance on outdoor and indoor scenes [8]. Multi-view stereo architectures designed for broad generalisation, such as MVSAnywhere, further extend this line of work by combining monocular and multi-view cues to enable zero-shot depth estimation across diverse domains and scales [9]. These developments motivate the modular pipeline adopted in this work, where different reconstruction back ends can be integrated and benchmarked against the specific challenges posed by organ and church interior imagery. 2.2 LLM-Assisted Georeferencing Georeferencing textual and visual material at scale is a long-standing problem in the digital humanities and cultural heritage domains, where historical place names, ambiguous toponyms, and incomplete descriptions frequently hinder automatic localisation. Recent advances in large language models (LLMs) have opened up new possibilities for combining probabilistic language understanding with structured queries to geospatial services, enabling more robust disambiguation of place references and the extraction of richer spatial context. In contrast to purely rule-based gazetteer matching, LLM-assisted approaches can incorporate broader context such as denominational affiliations, historical periods, or liturgical language when resolving church and organ locations. Recent surveys in GeoAI further highlight the potential of combining Large Language Models with Knowledge Graphs (KG), creating a retrieval-augmented generation (RAG) framework that grounds the probabilistic outputs of the LLM in verified geospatial ontologies, thereby reducing hallucination rates in toponym resolution [10]. Emerging work on LLM-guided entity linking in digital humanities pipelines further emphasises transparent provenance tracking and the explicit encoding of uncertainty rather than forcing overconfident point estimates [11]. 2.3 Heritage Databases The proposed pipeline operates in a broader landscape of cultural heritage aggregation initiatives that provide heterogeneous, but increasingly interoperable, access to digital surrogates and associated metadata. Europeana, for example, federates a large number of institutional repositories across Europe and exposes them via standardised protocols and schemas, thereby enabling cross-collection search and reuse of images, 3D models, and audiovisual materials [12]. Similarly, Smithsonian Open Access provides a substantial corpus of digitised artefacts and media under permissive licences (including CC0), which can serve as both training and evaluation data for reconstruction and analysis workflows in cultural heritage [13]. A critical enabler for this interoperability is the International Image Interoperability Framework (IIIF), which decouples image serving from presentation. IIIF APIs allow researchers to annotate, compare, and reconstruct assets across disparate repositories without data duplication, aligning with the FAIR principles (Findable, Accessible, Interoperable, Reusable) that now govern sustainable heritage data management [14, 15]. In the context of 3D content, infrastructures such as 3DRepo offer multi-user, web-based environments for storing, versioning, and visualising complex models, including architectural and archaeological datasets [16]. However, these platforms typically focus on geometric data and project-level metadata, with less emphasis on fine-grained provenance tracking for the underlying image sources and the intermediate processing stages. The MODAVIS Data database is designed to complement such repositories by modelling entities, events, and media at a more granular level, enabling intermodal links between images, documents, reconstructions, and acoustical models, while retaining detailed process and data provenance. 2 Large-Scale 3D Reconstruction of Pipe Organs Working Paper [v1] 2.4 Acoustical Heritage Modeling Acoustical modelling of heritage spaces has received growing attention in recent years, reflecting the importance of sound as an integral component of cultural experience in churches, concert halls, and other sacral or civic environments. Classical approaches are primarily based on geometrical acoustics, where ray-tracing or image-source methods are applied to polygonal models that approximate the architectural structure of the site. This has enabled interactive auralisation and in-situ exploration of historic sound fields, as demonstrated in research on location-based ”Virtual Acoustic Objects” for sacral environments [17]. However, geometric acoustics rely on high-frequency approximations that may insufficiently model diffraction and low-frequency resonance — phenomena critical to the sound of pipe organs. Consequently, wave-based methods such as Finite Difference Time Domain (FDTD) simulations are increasingly paired with geometric models to capture the full spectral response of heritage sites, provided the underlying mesh topology is watertight and free of non-manifold artifacts [18]. In parallel, researchers have investigated workflows to derive geometries for simulation directly from photogrammetric data. Early implementations of this approach in auditorium acoustics highlighted specific challenges regarding the accuracy of the generated meshes [19]. Llorca-Bofí et al. noted that while photogrammetry offers a user-friendly method to obtain spatial information, the resulting models must be ”closed” and contain material properties to be useful for acoustic simulation [19]. Their work on theatre spaces demonstrated that complex architectural features, such as balconies or hidden seating areas, often result in sparse data and mesh inaccuracies due to occlusion or insufficient lighting [19] .To mitigate this, high photographic overlap and carefully planned capture paths are required to avoid missing information in the dense point cloud [19]. More recently, the Photogrammetric-Acoustical Modeling Toolkit (PAMT) has been proposed to formalize this pipeline [20]. Ukolov introduces a parameterizable workflow that utilizes point reprojections to map local 2D coordinates from photographs to global 3D point identities, enabling the accurate remodeling of acoustical surfaces [20]. This methodology is designed specifically to support the Virtual Acoustic Object standard for the multimodal digitization of musical instruments—specifically pipe organs—and their coupling with the surrounding sacral architecture [20]. Unlike purely visual reconstructions, this approach employs neural networks for material classification and object segmentation to assign frequency-dependent absorption coefficients to the reconstructed surfaces [20]. Moving beyond explicit geometric modeling, joint audio-visual representations such as NeRAF have been introduced to learn coupled radiance and acoustic fields directly from collections of images and audio recordings [21]. NeRAF synthesizes both novel visual viewpoints and spatialized room impulse responses (RIRs) by conditioning the acoustic field on 3D scene geometric and appearance priors derived from the radiance field [21]. By using a grid sampler to query the NeRF, the system captures scene features that influence sound propagation, such as geometry and materials, without requiring explicit mesh annotations [21]. This suggests a future in which geometrical acoustics, image-based reconstruction, and neural field approaches are combined into hybrid pipelines for the documentation and experiential restitution of complex heritage spaces [20, 21]. 3 Data Sources Data acquisition was conducted in strict adherence to the MODAVIS framework standards, ensuring that all visual assets are integrally linked to their corresponding organ entity and provenance metadata. The aggregation pipeline automatically assesses image quality and enforces privacy compliance by detecting and masking personally identifiable information (PII) prior to downstream processing. 3 Large-Scale 3D Reconstruction of Pipe Organs Working Paper [v1] 3.1 Pipe Organ Databases A survey of the domain reveals a lack of open, API-accessible databases for pipe organs. Existing datasets are often distributed on physical media or via closed proprietary systems that do not adhere to FAIR principles. Consequently, this work aggregated web-based representations of these inventories, parsing and normalizing their heterogeneous structures into the unified MODAVIS Data schema. This approach facilitates intermodal linking while preserving granular provenance, ensuring that original authorship and source credits are maintained throughout the reconstruction pipeline. These provenance features facilitate the tracking of images from their respective sources, thereby enabling the identification of the creators or accredited sources of these images. 3.2 Documents Domain-specific literature and digitized archives were processed to identify organ entities and extract embedded visual content. Where document analysis identified photographic figures meeting quality thresholds, these images were extracted, linked to the entity, and ingested into the reconstruction pipeline. 3.3 Online Forums Online pipe organ communities represent a rich and persisting, yet underutilized, repository of visual documentation and technical discourse. The aggregation module targeted specific online forums to acquire high-quality photographs and video footage. While current efforts focus on German-language platforms, the architecture supports the extension to international forums. 3.4 Video Platforms Video platforms provide high-density visual data. The pipeline treats video streams as sequences of potential training images, subjecting them to frame extraction, blur detection, and the same privacy-preserving anonymization protocols applied to static imagery. Videos linked to entities in the test set were queried, identified, and processed to extract dense frame sequences. This continuous visual data provides the granular parallax and multi-view consistency required to maximize the inference performance of feed-forward architectures such as VGGT and π3, thereby bridging the coverage gaps inherent in sparse photography. 3.5 Other Web Resources To identify additional digital sources containing images of organs accompanied by reliable entity references, a thorough investigation was conducted. The websites of religious institutions and higher-level organizations were found to be particularly valuable resources in this regard. However, this necessitates an entity-specific web search, which was conducted using autonomous agents during the aggregation of MODAVIS Data due to the quantity involved. Following the identification of a photographic image, the image was aggregated and linked with the corresponding provenance and contextual data encoded in the source. 3.6 Geographical Databases MODAVIS Data incorporates a dedicated geographical database that captures the spatial context of pipe organ entities, associated persons, organisations, and provenance tracks. This database was built up cumulatively with each entity identification process, where structured queries were generated by LLMs, based on textual location context, and the responses evaluated within a multi-layer fallback process. Location information is derived from and cross-validated against established geospatial services, in particular GeoNames, Nominatim, and OpenStreetMap (OSM), and stored together with canonical coordinates, administrative hierarchies, historical place names, and names in other languages and scripts. The resulting records constitute a structured, 4 Large-Scale 3D Reconstruction of Pipe Organs Working Paper [v1] provenance-aware reference layer that supports the consistent identification, normalisation, and reuse of location data across the broader MODAVIS Data ecosystem. Beyond its role as an authority file for place information, the geographical database underpins a range of spatial and provenance-aware analyses within MODAVIS Data. The unified representation of locations enables spatial queries (for example, retrieval of all entities within a given radius, municipality, or historical administrative unit), supports the visualisation of organ distributions and movements of builders, and facilitates intermodal linking between images, documents, and other media via shared spatial references. By coupling location records with explicit provenance tracks, every geocoding decision remains traceable back to its textual evidence, the sequence of queried services, and the LLM-generated hypotheses, thereby ensuring both methodological transparency and reproducibility of the resulting spatial annotations. 4 Pipeline Architecture The pipeline’s architecture is highly modular to integrate most recent advancements in a specific domain. This work demonstrates this by substituting the reconstruction algorithm to evaluate the most performant for challenging interior objects like pipe organs. 4.1 Image Retrieval and Preprocessing All visual materials associated with the pipe organ entities are centrally referenced within MODAVIS Data, where they are linked to their corresponding organ records and provenance information. For each reconstruction task, the required images are retrieved via these links and downloaded into a temporary working environment, ensuring that all processing steps remain reproducible and traceable. This approach separates long-term archival storage from transient computational use, while preserving the explicit connections between images, their sources, and the higher-level entities to which they belong. Upon retrieval, image metadata are extracted using version 0.1.1 of the Heritage Data Processor (HDP) Component Image Metadata Extractor 1 . This component normalises and enriches technical metadata, including capture time, camera parameters, resolution, and, where available, geospatial information. The resulting metadata records are written back to MODAVIS Data and linked to the corresponding image entities, thereby providing a consistent basis for downstream analysis, quality assessment, and cross-modal alignment. To comply with data protection regulations and ethical requirements, all images undergo an automated person-detection and anonymisation step before being passed to the reconstruction algorithms. Person identification is currently performed using YOLOv10, with a planned migration to Ultralytics YOLO26 [22] as soon as it becomes available. For each detected bounding box, person segmentation is carried out using SAM2, from which binary masks are derived and applied to obscure persons within the image; this procedure is confined to the detection regions to minimise interference with relevant scene content. The corresponding HDP Component for person detection and segmentation is under active development 2 and will be integrated into the standard pipeline as it matures. Following anonymisation, the images are subjected to a series of quality assessments designed to identify and filter out data that would be detrimental to 3D reconstruction. To quantify these factors without reference images, the pipeline employs No-Reference Image Quality Assessment (NR-IQA) metrics, specifically the Naturalness Image Quality Evaluator (NIQE) and BRISQUE, to score perceptual degradation [23, 24]. These analyses encompass checks for excessive motion blur, severe compression artefacts, and exposure inconsistencies. Images that fall below defined thresholds are either excluded from further processing or marked with reduced confidence, depending on the severity of the detected issues. This systematic quality control step increases the robustness of the pipeline and reduces the computational cost associated with 1https://zenodo.org/records/16944352, doi: 10.5281/zenodo.16944352 2https://zenodo.org/communities/hdp-components/ 5 Large-Scale 3D Reconstruction of Pipe Organs Working Paper [v1] processing unusable material. In the final stage of preprocessing, image matching algorithms are employed to determine which image pairs and sets exhibit sufficient mutual overlap and feature correspondence to support multi-source reconstruction. Matching scores are computed and compared against empirically defined thresholds, ensuring that only those image sets with adequate geometric and photometric compatibility are propagated into the 3D reconstruction stage. This selective pairing mechanism is particularly crucial in heterogeneous, multi-source settings such as the present one, where images may originate from different cameras, time periods, and acquisition conditions. By enforcing compatibility at the level of image sets, the pipeline enhances reconstruction stability while preserving the flexibility to incorporate diverse sources. For the geometric verification of these sets, the pipeline integrates deep-learning-based matchers such as LightGlue, which effectively filter outliers in challenging image pairs exhibiting wide baselines or repetitive textures—common features in organ pipe fields [25]. The resulting correspondence graphs are pruned using MAGSAC++, a robust estimator that offers superior stability over standard RANSAC by marginalising over the noise scale, ensuring that only geometrically consistent clusters are passed to the reconstruction back end [26]. 4.1.1 Algorithms for Image Matching Evaluation In the current implementation, image matching builds on a combination of local feature extraction, approximate nearest-neighbour search, and robust geometric verification. Local features are derived using state-of-the-art keypoint detectors and descriptors, which are subsequently matched across image pairs and filtered via ratio tests and bidirectional consistency checks to suppress ambiguous correspondences [1]. The filtered matches are then subjected to epipolar verification using modern robust estimators such as MAGSAC++, which offers superior stability over standard RANSAC by marginalizing over the noise scale [26]. This is calculated via the essential or fundamental matrix, followed by optional local bundle adjustment to refine relative camera poses where sufficient parallax is available [1]. To quantify and select suitable image constellations for multi-source reconstruction, the pipeline computes pairwise matching scores based on the number and spatial distribution of inlier correspondences. However, exhaustive pairwise matching is computationally prohibitive for large unordered collections. Furthermore, the specific geometry of pipe organs—characterized by repetitive vertical ranks and symmetrical casework—poses a risk of ambiguous loop closures, where similar but distinct features are incorrectly matched. To counteract this, the pipeline can enforce loop consistency constraints during graph construction, rejecting edges that generate geometrically inconsistent triplets, thereby preventing the ’folded’ or superimposed structures common in reconstructions of symmetric heritage objects [27]. This strategy is particularly important in the present setting, where images stem from a wide range of cameras, time periods, and acquisition conditions; enforcing minimum connectivity and geometric consistency significantly reduces failure modes and improves the stability of downstream reconstruction. 4.2 3D Reconstruction The 3D reconstruction stage of the pipeline has been designed to accommodate multiple algorithmic back ends, reflecting the rapid evolution of the field and the heterogeneous nature of the available image data. In addition to a classical SfM+MVS baseline implemented via COLMAP, several recent learning-based methods have been integrated and evaluated on the organ and church interior test cases, including DUSt3R, VGGT, π3, and MVSAnywhere [1, 6–9]. DUSt3R provides a geometry-centric baseline that predicts dense point maps and camera poses directly from image pairs or small sets, and has shown strong generalisation to real-world scenes with limited or unstructured coverage [8]. VGGT extends this concept by inferring all key 3D attributes of a scene—camera parameters, depth maps, and dense point clouds—in a single transformer-based forward pass, which is attractive for large-scale processing due to its speed and simplicity [6]. The permutation-equivariant design 6 Large-Scale 3D Reconstruction of Pipe Organs Working Paper [v1] of π3 further relaxes assumptions about reference views and input ordering, allowing more flexible handling of image collections that exhibit varying degrees of redundancy and overlap [7]. MVSAnywhere complements these approaches with a general-purpose multi-view stereo architecture explicitly designed for zero-shot application across diverse domains and depth ranges [9]. By combining monocular cues with an adaptive cost volume over multi-view features, MVSAnywhere can robustly estimate depth even in the presence of scale variation and textureless regions, which are common in large interior spaces such as churches [9]. For the purposes of this study, these methods are treated as interchangeable modules behind a common interface: given a curated and matched image set, the selected back end produces a dense point cloud or depth map ensemble, which is then converted into a mesh or a Gaussian Splatting representation suitable for visualisation and downstream acoustical analysis [3]. For acoustical simulation specifically, point clouds are meshed using Screened Poisson Surface Reconstruction to approximate watertight, manifold geometries required by boundary element methods, filling holes caused by occlusion while preserving high-frequency architectural details [28]. 4.3 Georeferencing Following reconstruction, each 3D model must be situated within a consistent spatial reference frame. In the present workflow, georeferencing is performed by aligning the reconstructed geometry with canonical location records from the MODAVIS geographical database. 4.3.1 LLM-Driven Location Resolution To populate this database, the pipeline utilizes a dedicated HDP Component that leverages Large Language Models (LLMs) to resolve unstructured textual descriptions. The LLM generates structured, service-specific queries (e.g. for GeoNames, Nominatim, or OpenStreetMap) based on input metadata, and orchestrates a multilayered fallback strategy when initial attempts at geocoding are inconclusive. The resulting georeferencing process not only yields canonical coordinates and administrative hierarchies, but also records the sequence of hypotheses and service responses, thus providing a reproducible trail of the reasoning steps involved. For the test dataset, entity matching between reconstruction outputs and MODAVIS Data is carried out using a combination of organ-specific identifiers (where available) and this LLM-assisted pipeline. In cases where multiple candidate locations exist — for example, when several churches share the same dedication — the system records all plausible matches together with confidence scores and justifications, allowing researchers to inspect and, if necessary, correct individual assignments. [Figure 1: Schematic of the LLM-Georeferencing data flow. Detailed architectural diagram to be included in version 2.] Figure 1: The proposed HDP Georeferencing workflow: textual metadata is processed by the LLM to generate structured queries for GeoNames/OSM. 7 Large-Scale 3D Reconstruction of Pipe Organs Working Paper [v1] 4.3.2 Geometric Alignment When suitable reference geometry is available, the pipeline supports an additional refinement step in which reconstructed models are rigidly aligned to building footprints or outline models derived from OpenStreetMap or other sources. This can help correct residual scale or rotation ambiguities, particularly in cases where camera intrinsics are uncertain or where the image network covers only a portion of the building interior. The current work employs this option selectively, as automatic alignment in complex, multi-room churches remains challenging and is deferred to future research on view-based and VLM-assisted (Visual Language Model) alignment strategies. 4.4 Database Integration A central design principle of the pipeline is that all processing steps interact with MODAVIS Data rather than with ad hoc file system structures. Input images and their metadata are retrieved from the database using organ-level identifiers and provenance links, and all intermediate and final products are written back as new entities or events that extend existing provenance chains. This approach ensures that reconstruction results are not isolated artefacts, but integral components of a broader, intermodal data graph that also encompasses documents, forum posts, videos, and acoustical models. For each reconstruction run, the pipeline records configuration parameters, software versions, and references to the specific HDP Components invoked, thereby enabling reproducibility and comparative evaluation across different algorithmic back ends and parameter settings. The resulting model entities are annotated with links to their source images, the associated organ and building entities, and the georeferenced location records, as well as with derived quantitative descriptors such as point counts, mesh statistics, and basic quality indicators. This rich annotation supports subsequent analyses, for example when correlating reconstruction quality with acquisition conditions, organ typologies, or spatial contexts. To facilitate dissemination and long-term preservation, the integration layer also prepares metadata packages for deposition in external repositories such as Zenodo. Here, MODAVIS Data acts as the authoritative source for related identifiers: the deposited records can reference not only the 3D models themselves but also associated images, documents, and geospatial entities via persistent identifiers, enabling cross-repository navigation and reuse. The HDP infrastructure and agentic metadata filling tools are used to automate as much of this process as possible while still allowing expert oversight and correction. 4.5 Storage and Access From a storage perspective, the pipeline distinguishes between transient computational artefacts and long-term archival products. Intermediate files such as temporary point clouds, cache files, and diagnostic visualisations are kept in a local working environment only for the duration of processing and, where appropriate, summarised as derived statistics before being discarded as paradata extensions. In contrast, final reconstructions, together with their associated provenance and metadata, are persisted either in MODAVIS-managed storage or in external repositories such as Zenodo, depending on licensing conditions and anticipated reuse scenarios. Access to the resulting models is mediated by MODAVIS Data, which provides both human-facing interfaces and machine-accessible APIs for querying and retrieving entities based on organ attributes, spatial criteria, provenance constraints, or reconstruction parameters. For external repositories, the system relies on related identifiers and rich metadata to maintain logical connections even when the actual files are stored elsewhere. This design supports a range of usage scenarios, from interactive exploration of individual organs to large-scale analyses of spatial distributions, reconstruction quality, or acoustical properties across the corpus. 8