TL;DR
This paper tackles the challenge of loading large Industry Foundation Classes (IFC) models in web browsers, which often leads to performance issues and crashes. They propose a novel "spatial semantic partitioning" (SSP) method that dynamically loads only the necessary parts of an IFC model based on user interaction and the model's inherent semantic structure, significantly improving loading speed and reducing memory consumption.
What Problem Does This Paper Address?
Building Information Modeling (BIM) is increasingly used in Architecture, Engineering, and Construction (AEC) for its rich geometric and semantic data. Industry Foundation Classes (IFC) is a crucial open standard for BIM data interoperability. However, as BIM applications move to the web, loading large IFC models becomes a major bottleneck. These models can be massive, consuming excessive browser memory and often leading to crashes, thus limiting their practical use in web-based scenarios. Existing solutions often involve loading the entire model or using generic spatial partitioning methods that ignore the rich semantic information embedded within IFC files. This paper aims to bridge this gap by enabling fast, dynamic loading of large IFC models on the web, enhancing user experience and enabling new web-based BIM applications.
Methods
The core of the proposed solution is a Spatial Semantic Partitioning (SSP) method. This method operates in two main stages:
- Server-side Pre-computation:
- External Component Extraction: Instead of relying solely on bounding boxes (which can miss components like windows embedded in walls), this method uses a node classification approach. It divides the model's bounding box into smaller "cube nodes" and checks for intersections with components. This ensures accurate identification of external elements.
- Story Information Extraction: IFC models inherently contain hierarchical information about building stories. The algorithm leverages
IfcBuildingandIfcBuildingStoryentities to extract and organize components by their respective stories. This creates a semantic structure based on the building's layout. - Spatial Index Table Creation: A spatial index is built to facilitate efficient local navigation. This index maps components to the spatial partitions (cube nodes) they belong to.
- Client-side Dynamic Loading:
- User Interaction-driven Loading: When a user interacts with the model (e.g., zooms in on a specific area), the system identifies the relevant components based on the pre-computed SSP and the user's focus.
- Adaptive Network Transmission: To handle varying network conditions, an adaptive transmission algorithm is employed. It dynamically adjusts the size of data blocks being sent based on current network speed, ensuring a smoother experience.
- WebGL Implementation: The entire system is implemented using Web Graphics Library (WebGL), a standard web API for 3D graphics, ensuring it runs in modern browsers without requiring any plugins.
Datasets
The paper evaluates its method using three large IFC models (M1, M2, M3) from the industry foundation classes. These models are described by their file size (MB) and the number of components. No new datasets were introduced.
Key Insights and Ideas
- Leveraging IFC Semantics is Key: Unlike generic 3D spatial partitioning, this method's strength lies in exploiting the rich, built-in semantic information of IFC files (like story structure and component types). This semantic understanding is crucial for effective partitioning.
- External Components First: Prioritizing the loading of external components provides users with a quick, recognizable overview of the model, significantly improving the initial browsing experience.
- Node Classification for Robustness: The proposed external component extraction method, based on node classification rather than just bounding boxes, is more robust and avoids missing components that might be obscured by larger bounding boxes.
- Dynamic Loading for Performance: By only loading what the user is actively viewing or interacting with, the system drastically reduces memory footprint and improves loading times, making large IFC models manageable in a web environment.
Key Results
- Memory Reduction: The proposed method reduced memory consumption by approximately 70% compared to conventional methods that load the entire IFC model. For example, one test model used 4.5 GB with the conventional method versus 1.7 GB with the proposed method.
- Faster Loading: While loading 100% of the data took a similar amount of time for both methods, the proposed method provided a good overview of the model by loading only 35% of the data in significantly less time (e.g., 12 seconds for M1 vs. 41 seconds for full load).
- Improved Display Effect: The node classification-based external component extraction accurately identified components like windows and doors, which were missed by bounding-box-based methods.
- Smoother Operation: The frame rate (FPS) for interactive operations was significantly higher with the proposed method, especially when displaying only external components, leading to a more fluent user experience. For example, FPS increased from 3 to 20 FPS for one model.
Limitations and Future Work
- Robustness of External Component Extraction: The authors acknowledge that the current external component extraction algorithm is not fully robust for non-closed BIM models, such as those for mechanical, electrical, and plumbing (MEP) or structural systems, which might not explicitly include architectural context.
- MEP and Structural Models: Distinguishing external from internal components in MEP and structural models remains a challenge.
- Memory Management: Further improvements in memory management strategies are suggested for future work to enhance dynamic loading performance.
- Future Directions: The authors plan to explore more effective algorithms for external component extraction, particularly for MEP and structural models, and investigate advanced memory management techniques.
Summary updated