FaçAID: A Transformer Model for Neuro-Symbolic Facade Reconstruction

Aleksander Plocharski,Jan Swidzinski,Joanna Porter-Sobieraj,Przemyslaw Musialski
2024-09-13
Abstract:We introduce a neuro-symbolic transformer-based model that converts flat, segmented facade structures into procedural definitions using a custom-designed split grammar. To facilitate this, we first develop a semi-complex split grammar tailored for architectural facades and then generate a dataset comprising of facades alongside their corresponding procedural representations. This dataset is used to train our transformer model to convert segmented, flat facades into the procedural language of our grammar. During inference, the model applies this learned transformation to new facade segmentations, providing a procedural representation that users can adjust to generate varied facade designs. This method not only automates the conversion of static facade images into dynamic, editable procedural formats but also enhances the design flexibility, allowing for easy modifications.
Graphics,Artificial Intelligence,Computer Vision and Pattern Recognition,Machine Learning,Neural and Evolutionary Computing
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: how to automatically convert the building facade images with planar segmentation into editable procedural definitions, thereby simplifying and automating the design and modification process of building facades. Specifically, although the existing building facade modeling methods can efficiently generate complex urban structures and environments, creating syntax rules and their parameters requires professional knowledge and is time - consuming, which poses certain obstacles to non - professional users. In addition, although the reverse procedural modeling methods can extract procedural rules from existing models or images, these methods have problems such as high complexity, error - proneness, and limited flexibility and variability. For this reason, the paper proposes a Transformer - based neuro - symbolic model, which combines the learning ability of neural networks and the symbolic reasoning ability of procedural models to achieve the following goals: 1. **Automatic conversion**: Automatically convert the building facade structures with planar segmentation into editable procedural definitions. 2. **Custom - made split grammar**: A split grammar specifically for building facades has been designed. 3. **Dataset construction**: A dataset containing building facades and their corresponding procedural definitions has been created as the basis for training and validation. 4. **Practical application**: Demonstrated the practicality and flexibility of the model in generating diverse building facade designs. Through this method, users can not only automatically generate procedural definitions, but also easily adjust and generate new facade designs, greatly improving the flexibility and creativity of design. ### Main contributions - Proposed a Transformer - based neuro - symbolic model, which realizes the automatic conversion from segmented facade images to editable procedural definitions. - Designed a split grammar specifically for building facades. - Constructed a dataset containing facade structures and their procedural definitions. - Demonstrated the practical application of the model in generating diverse building facade designs, proving its practicality and flexibility. ### Method overview - **Data generation**: Developed a generation system based on split grammar to generate diverse building facades and their corresponding procedural definitions. - **Model architecture**: Used the Transformer model for neuro - symbolic learning, combined with the procedural engine and the optimization module to realize the conversion from segmented facades to procedural definitions. - **Optimization and adjustment**: By minimizing the mean square error (MSE), optimized the size parameters in the generated procedural definitions to match the input facade images. ### Experimental results - The model can accurately reconstruct the split grammar in most cases and has a certain robustness to input noise. - The optimization module can highly match the generated structures with the target segmentation, and the matching degree of most facades reaches more than 80%. Through these innovations, this paper provides an efficient, flexible and easy - to - use tool for the design and modification of building facades.