IfcVoxNet 3D segmentation and classification of voxelized IFC building models with Deep Learning

Transcript

1. IfcVoxNet 3D segmentation and classification of voxelized IFC building models with Deep Learning Johan Luttun, johan@aecgeeks.com Thomas Krijnen, thomas@aecgeeks.com Presenter: Artur Tomczak CIB W78 2024 Marrakesh, Morocco
2. Context: Examples of misclassifications in BIM models Elements classified as IfcWall in the Duplex_A model. Source: Krijnen, T., & Tamke, M. (2015). Assessing implicit knowledge in BIM models with machine learning. In Modelling Behaviour: Design Modelling Symposium 2015 (pp. 397-406). Springer International Publishing.
3. Context: Advanced automated use cases requiring accurate classifications Quantity takeoffs Building permit checks Simulations …
4. Context: Enabling more granular classifications Regulations on the height of door handles, automatic doors closers, … IfcDoor Hinge Handle 0.85 < d < 0.95
5. The importance of geometric context Ambiguity: wall, beam, column, railing, or just wall? Problem definition: BIM elements have little discriminating geometric features, but require context to interpret their role
6. Modeler, software, LOD differences Problem definition: Geometry and topology of elements varies based on modeler, software and LOD.
7. Meanwhile in 2D image recognition “Chihuahua or muffin” (source unknown) How can we benefit from advancements in convolutional neural networks with Voxels as a “3D pixels”?
8. Other uses for voxelization https://github.com/opensourceBIM/ voxelization_toolkit
9. Network architecture 3D U-Net architecture Çiçek, Özgün, et al. "3D U-Net: learning dense volumetric segmentation from sparse annotation." Medical Image Computing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. Springer International Publishing, 2016.
10. Approach 1. Voxelize IFC models. 1 voxel belongs to 1 class (order of priority set) 2. Train/test splitting 3. Create 64x64x64 3D patches from individual instances bbox centers 4. Train with weighted Cross-entropy loss 5. Use the model in inference
11. Dataset
12. Dataset
13. Results presented graphically Confusion at the boundaries between elements. Network ‘expects’ window to come with window sill. wall window slab stair beam railing error
14. Results presented graphically Staircase accurately predicted, but the ‘social stairs’ is not recognized as a stair (it most likely also isn’t). wall window slab stair beam railing error
15. Results presented graphically Confusion between whether the final step of the staircase is part of the stair or part of the floor slab. wall window slab stair beam railing error
16. Results confusion table Per-voxel Model: FZK-Haus
17. Results confusion table Per-element Model: Molio
18. Results confusion table Per-element Model: FZK-Haus
19. Conclusions and next steps - Promising results for further research and applications - Test instance segmentation (door knob, door closers) - Find the ideal patch size (#voxels) and voxel size (cm) - Add material characteristics to voxels in addition to class - Use multiclass per voxel - Anomaly detection
20. Anomaly detection boredpanda.com
21. IfcVoxNet 3D segmentation and classification of voxelized IFC building models with Deep Learning Johan Luttun, johan@aecgeeks.com Thomas Krijnen, thomas@aecgeeks.com Presenter: Artur Tomczak CIB W78 2024 Marrakesh, Morocco