Converting CAD Data to Graphs for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for converting unstructured technical drawing or CAD data into map data are inefficient, leading to loss of detail and increased computational effort due to rasterization, and struggle with inconsistency in data representation, which hampers reliable object detection and map creation.
Innovation Solution
A computer-implemented method that converts unstructured map data into a graph data representation, where geometric entities are represented as nodes and edges, allowing for a more structured format that preserves geometric properties and enables efficient object detection using graph neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If technical drawing data is rasterized into a 2D image for object detection, then the data can be processed by neural networks, but geometric details and information such as precise coordinates, resolution, and attributes are lost
Solution Approach 1:
The patent introduces an intermediary conversion process that transforms unstructured technical drawing data into a structured graph data representation before feeding it to neural networks. This graph representation serves as a mediator that preserves geometric details, coordinates, and attributes while making the data suitable for automated processing by graph neural networks, thereby avoiding the information loss inherent in rasterization.
2Extent of automation
If rasterization is used to convert technical drawing data to images, then object detection can be performed, but computational effort and storage requirements increase significantly
Solution Approach 1:
The patent fundamentally changes the data representation parameter from raster/image format to graph data format. This parameter change reduces computational complexity because graph representations maintain the structural and geometric information in a more compact and processing-efficient form, allowing graph neural networks to operate with lower computational effort and storage requirements compared to high-resolution images.
3Extent of automation
If conventional neural networks are used on rasterized technical drawings, then object detection is possible, but the convolutional nature assumes neighboring points have similar attributes which is not true for technical drawing data
Solution Approach 1:
Instead of forcing technical drawing data into the conventional 2D image convolutional framework, the patent inverts the approach by adopting graph data structures and graph neural networks that are inherently suited for representing and processing the topological and geometric relationships in technical drawings. This inversion allows the model to naturally handle the attribute variability of neighboring points without relying on the similarity assumption that underpins conventional CNNs.
4Reliability
If manual or semi-automatic generation of map data from technical drawings is used, then object detection can be performed, but the process is time-consuming and requires professional review and correction
Solution Approach 1:
The patent enables the system to automatically process and convert unstructured technical drawing data into structured graph representations without requiring manual intervention or professional review. The graph neural network performs object detection and map data generation autonomously, allowing the system to serve itself by handling the entire workflow from raw technical drawings to structured map data without human involvement, thereby eliminating time losses associated with manual correction and standardization.
5Adaptability or versatility
If different datasets with inconsistent representation standards are processed, then comprehensive map coverage is achieved, but professional review and standardization are required before reliable automatic detection
Solution Approach 1:
The patent creates a universal graph data representation framework that can accommodate and process technical drawing data from multiple different datasets and sources with varying representation standards. The graph neural network is designed to handle diverse input formats and automatically adapt to different data structures, eliminating the need for separate processing pipelines or manual standardization for each data source, thereby reducing processing complexity while maintaining adaptability.
Data Source
AI summary
A computer-implemented method of converting map data. The method includes: obtaining unstructured map data according to a first data representation, the unstructured map data representing or including a number of geometric entities where the first data representation is a technical drawing representation or a CAD data representation, and converting the unstructured map data according to the first data representation to structured map data according to a second data representation, where the second data representation is a graph data representation.