Floorplan Generation from 3D Scans Using Deep Network Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for estimating room layouts from indoor scenes face challenges such as reliance on hand-engineered features, susceptibility to clutter, and limitations in processing complex indoor environments, leading to inaccurate and inefficient floorplan generation, especially in cluttered scenes with non-box-shaped topology.
Innovation Solution
A method involving deep networks trained on synthetic datasets to classify rooms and walls from 3D scans, using nested partitioning and semantic feature extraction to generate room and wall cluster labels, and a voting architecture for efficient floorplan estimation without constraints on the number of rooms or room size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional methods use hand-engineered features and vanishing point detection for floorplan generation, then the process is simpler to implement, but the accuracy and reliability deteriorate in cluttered scenes with complex topology
Solution Approach 1:
The patent replaces conventional hand-engineered feature extraction and vanishing point detection methods with deep learning-based automated feature extraction. The system uses convolutional neural networks to automatically learn and extract relevant features from images, eliminating the need for manual feature engineering and improving accuracy in cluttered scenes while maintaining implementation simplicity through end-to-end training.
2Adaptability or versatility
If deep networks are trained on synthetic datasets with complex indoor environments, then the accuracy and adaptability improve, but the training time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by training deep networks on synthetic datasets in advance before actual floorplan generation is needed. The pre-trained models capture general patterns of indoor environments, allowing the system to quickly adapt to specific scenes without requiring extensive retraining. This approach balances adaptability with computational efficiency by performing the heavy lifting beforehand.
3Adaptability or versatility
If the system processes unlimited number of rooms and varying room sizes without constraints, then the versatility and adaptability improve, but the processing time and computational complexity increase
Solution Approach 1:
The patent applies segmentation by dividing the floorplan generation process into independent steps: image processing, feature extraction, room detection, and floorplan construction. Each step processes local features and can handle varying numbers of rooms and sizes independently, allowing the system to maintain versatility while optimizing processing speed through modular computation.
Data Source
AI summary
Methods, systems, and wearable extended reality devices for generating a floorplan of an indoor scene are provided. A room classification of a room and a wall classification of a wall for the room may be determined from an input image of the indoor scene. A floorplan may be determined based at least in part upon the room classification and the wall classification without constraining a total number of rooms in the indoor scene or a size of the room.