Vector Map Generation Using Tokenized Annotation Operations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for extracting vector maps from imagery are laborious and lack flexibility, and vertex-based models struggle with generating complex geometric primitives and spatial constraints.
Innovation Solution
A machine learning model, particularly autoregressive generative models with attention mechanisms, directly converts raster imagery into vector maps by learning sequences of annotation operations, capturing geometric models and spatial constraints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation methods are used to extract vector maps from imagery, then accuracy can be maintained, but the process becomes laborious and time-consuming
Solution Approach 1:
The patent replaces manual mechanical annotation processes with an automated machine learning system. The ML model processes imagery and directly generates vector map features, eliminating the need for human annotators to manually draw polygons and lines while maintaining extraction accuracy through learned patterns from training data.
Solution Approach 2:
The system enables self-service annotation by training the machine learning model on annotated training data, allowing it to autonomously perform vector map extraction without continuous human intervention. Once trained, the model independently processes new imagery and generates accurate vector features automatically.
2Extent of automation
If vertex-based models are used to convert raster images to vector maps, then automation is achieved, but the models struggle with generating complex geometric primitives and spatial constraints
Solution Approach 1:
The patent changes the output parameters of the automated model from simple vertex coordinates to structured geometric primitives with explicit spatial constraints. The model generates features with parameters such as line orientations, polygon shapes, and relative position relationships, enabling it to handle complex geometric structures while maintaining automation.
Solution Approach 2:
The patent segments the vector map generation process into distinct geometric primitives (lines, polygons, curves) with specific spatial constraint relationships. This segmentation allows the model to handle complex structures by composing them from simpler, well-defined geometric elements with explicit spatial relationships.
3Ease of manufacture
If raster labels are used to represent semantic information, then extraction is straightforward, but the data lacks scalability and editability compared to vector data
Solution Approach 1:
The patent substitutes raster label generation with direct vector feature extraction. Instead of producing pixel-based semantic masks that require subsequent conversion, the system directly outputs editable vector geometric primitives with spatial constraints, combining the simplicity of automated extraction with the versatility of vector data formats.
Data Source
AI summary
Methods and systems for generating vector maps representing features depicted in imagery are provided. An example method involves obtaining remote imagery that depicts a feature, applying a machine learning model to the remote imagery to extract a geometric model of the feature encoded as a tokenized representation of a sequence of annotation operations, and interpreting the tokenized representation as a vector map representing the geometric model of the feature.