TL;DR
This paper introduces "A-Scan2BIM," an assistive system that helps architects convert 3D scan data into Building Information Models (BIM) by predicting sequences of modeling operations. They also present a novel dataset and evaluation metric to better assess the "naturalness" of these predicted sequences, demonstrating superior performance compared to baselines.
What Problem Does This Paper Address?
Creating Building Information Models (BIM) from 3D scans of existing buildings, a process known as "Scan-to-BIM," is crucial for various industries like construction and renovation. However, this process is notoriously labor-intensive and time-consuming, often requiring many hours of manual work from professional architects, even for a single floor. While automatic Scan-to-BIM tools exist, they often suffer from poor performance or integration issues with architects' workflows, leading them to be underutilized. This paper aims to bridge this gap by developing an assistive system that helps architects rather than replacing them, by intelligently suggesting the next modeling steps.
Methods
The core of A-Scan2BIM is a two-stage process:
- Candidate Wall Enumeration: This stage leverages a modified version of the state-of-the-art floorplan reconstruction system, HEAT. It takes raw sensor data and an existing BIM model to identify potential walls. This involves:
- Corner Enumeration: Using a Transformer network to detect corners.
- Wall Enumeration: Employing a modified deformable attention layer to better handle long edges.
- Wall Thickness Prediction: An additional MLP estimates wall thickness.
- Duplicate Removal: Filters out redundant wall candidates.
- Next Wall Prediction: This is the heart of the assistive system. An auto-regressive Transformer network takes the current BIM model state (including edit history) and the enumerated wall candidates. It then predicts a sequence of modeling operations (specifically, wall addition APIs for Autodesk Revit). The network learns to score candidate walls based on their cosine similarity in an embedding space with the last added wall, aiming to predict the most likely next step.
The system operates in two modes:
- Automatic Mode: Reconstructs the BIM model without user interaction by greedily selecting the highest-scoring candidate at each step.
- Assistive Mode: Offers wall auto-completions to the architect after each manual interaction, allowing them to accept or reject suggestions.
Datasets
The paper introduces the first building-scale Scan2BIM dataset. This dataset comprises:
- 89 hours of modeling processes.
- Data from 16 scenes across 11 buildings.
- Covers over 35,000 square meters.
- Includes sequences of model editing operations as APIs from Autodesk Revit, captured from professional architects.
This dataset is notable for its scale and the inclusion of actual human modeling sequences, which is crucial for training and evaluating assistive systems.
Key Insights and Ideas
- Assisting, Not Replacing: The fundamental insight is that the most effective approach for Scan-to-BIM is to augment architects' capabilities rather than trying to fully automate the process, which often fails to align with their intuitive workflows.
- Sequence Prediction for BIM: Modeling BIM elements is inherently sequential. Predicting this sequence of operations, rather than just the final geometry, is key to creating a natural and efficient workflow.
- "Naturalness" as a Metric: Traditional accuracy metrics are insufficient for evaluating modeling sequences. The introduction of a novel "order metric" inspired by FID score, which measures the "naturalness" of the reconstruction order, is a significant contribution. This metric better captures how well the predicted sequence aligns with human-like modeling behavior.
- Transformer for Sequential Operations: The Transformer architecture, known for its success in sequence modeling, proves effective in predicting the next wall addition operation, demonstrating its versatility beyond natural language processing.
Key Results
- Superior Order Metric Performance: The proposed A-Scan2BIM system significantly outperforms two baseline methods (heuristic and classifier) on the novel "naturalness" metric, indicating its ability to generate more architect-like modeling sequences.
- High Accuracy: The system also shows strong performance on standard reconstruction metrics.
- Impact of Deformable Attention: Modifying the deformable attention module with more reference points leads to improved precision, recall, and IoU scores for wall reconstruction.
- History Length Matters: Increasing the history length for the next wall prediction module leads to higher accuracy and lower entropy in predictions, confirming the hypothesis that more context improves performance.
Limitations and Future Work
The authors acknowledge that their system still exhibits failure cases in both reconstruction and order prediction, which need further consideration for real-world deployment. They pledge to continue development and make their data, code, and models publicly available to foster further research in this area. Future work could focus on improving the robustness of reconstructions and further refining the prediction of complex modeling sequences.
Summary updated