3D Model Generation from 2D Floor Plans Using Semantic Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for converting 2D architectural floor plans into 3D models are inefficient due to reliance on image-processing and pattern-recognition techniques, which lack generality and require extensive user intervention, leading to errors and lengthy processes.
Innovation Solution
A computer-implemented method using machine-learning, specifically a convolutional encoder-decoder neural network, for semantic segmentation of 2D floor plans to accurately and efficiently generate 3D models by converting semantic segmentation into a 2D model and then into a 3D model, reducing the need for user intervention and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If image-processing and pattern-recognition techniques are used to convert 2D floor plans into 3D models, then the conversion process can be automated, but the system lacks generality and requires extensive user intervention
Solution Approach 1:
The patent segments the floor plan image into multiple meaningful regions (walls, windows, doors, furniture, text elements) using line detection and clustering algorithms. This segmentation enables the system to handle diverse floor plan representations by identifying and processing different architectural elements separately, thus improving both automation and adaptability.
Solution Approach 2:
The patent employs machine learning models with learnable parameters that adapt to different floor plan styles and representations. By training the model on diverse datasets with varying line thicknesses, colors, and symbols, the system gains generality while maintaining automated processing capability across different architectural drawings.
2Measurement precision
If pattern recognition methods are used to identify floor plan symbols, then specific symbol types can be recognized, but the approach is constrained to a limited set of predefined symbols
Solution Approach 1:
The patent implements a universal line-based representation model that can recognize various architectural symbols (walls, windows, doors, furniture) through common geometric primitives. Instead of requiring specific predefined symbol patterns, the system uses universal line detection and spatial relationship analysis that adapts to different symbol types, thereby improving both precision and versatility.
Solution Approach 2:
The patent extracts geometric features and spatial relationships from the input floor plan and creates a standardized intermediate representation that captures the essential structure. This copied representation is then used for 3D model generation, allowing the system to handle diverse symbols by focusing on their geometric properties rather than specific visual patterns.
3Reliability
If multiple processing steps are used to clean and process 2D plans, then errors can be reduced, but errors and inconsistencies are accumulated from one step to another
Solution Approach 1:
The patent merges multiple error-prone processing steps into a unified end-to-end machine learning pipeline. Instead of sequentially applying preprocessing, line detection, symbol recognition, and 3D generation as separate steps that accumulate errors, the integrated model processes the input through all stages simultaneously, reducing error propagation while maintaining reliability.
Solution Approach 2:
The patent incorporates loss functions and optimization mechanisms that provide feedback during training to minimize errors in the final 3D model generation. The model learns to adjust its internal representations and processing decisions based on the quality of the output, thereby maintaining high reliability without requiring numerous intermediate correction steps.
4Measurement precision
If user intervention is required to correct problems in the original file, then accuracy can be improved, but the process becomes costly and very long (non-real-time)
Solution Approach 1:
The patent implements a self-correcting machine learning model that automatically handles errors and inconsistencies in the input floor plan without requiring user intervention. The model learns to robustly process imperfect inputs by training on diverse and noisy data, thereby maintaining high accuracy while achieving real-time processing speeds through automated error correction.
Solution Approach 2:
The patent performs preliminary learning and adaptation during the training phase, where the model learns to handle various error patterns and inconsistencies that may appear in input floor plans. This preliminary action equips the model with the ability to automatically correct errors during inference, eliminating the need for time-consuming user corrections while maintaining accuracy.
5Measurement precision
If a value approaching standard wall thickness is chosen for wall detection, then wall recognition can be improved, but different thicknesses exist depending on the construction domain and designers
Solution Approach 1:
The patent uses dynamic parameter adjustment where the model learns optimal wall thickness detection thresholds adaptively during training based on the specific construction domain and design conventions of the training data. This allows the system to maintain high detection accuracy while being versatile across different construction domains, as the parameters dynamically adapt to the characteristics of each domain rather than using fixed values.
Data Source
AI summary
The disclosure notably relates to a computer-implemented method for generating a 3D model representing a building. The method comprises providing a 2D floor plan representing a layout of the building. The method also comprises determining a semantic segmentation of the 2D floor plan. The method also comprises determining the 3D model based on the semantic segmentation. Such a method provides an improved solution for processing a 2D floor plan.