Template-Volume Labelling for Automated 3D Object Representations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The manual process of labeling 3D data for AI training is time-consuming and costly, and existing methods for generating labeled 3D data are laborious and expensive, especially when training AI in both 2D and 3D domains.
Innovation Solution
A method and system for generating labeled 3D data representations of real-world objects by obtaining pre-labelled 3D coordinates, applying template volumes, and labeling object volumes without manual intervention, using processing circuitry and sensors on unmanned vehicles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual image recognition is performed by a person to obtain labelled training data, then the AI algorithm can be trained with accurate labels, but the process becomes very expensive and time-consuming
Solution Approach 1:
The patent performs preliminary action by manually labelling only a small subset of 3D objects first, then uses this initial labelled data to train an AI model that automatically labels the remaining objects. This preliminary manual labelling provides the foundation for subsequent automated labelling, dramatically reducing total labelling time while maintaining accuracy.
Solution Approach 2:
The patent introduces an AI model as an intermediary between manual labelling and final labelled dataset creation. The AI model is trained on manually labelled data and then used to automatically label remaining objects, serving as a mediator that transfers labelling knowledge from human experts to the entire dataset without requiring direct human involvement in labelling every object.
2Quantity of substance
If a large amount of labelled 3D data is gathered for AI training, then the training can be conducted satisfactorily in the 3D domain, but the process becomes slow, laborious, and expensive
Solution Approach 1:
The patent performs preliminary action by manually labelling only a small subset of 3D objects first, then uses this initial labelled data to train an AI model that automatically labels the remaining objects. This preliminary manual labelling provides the foundation for subsequent automated labelling, dramatically reducing total labelling time while maintaining accuracy.
Solution Approach 2:
The patent implements self-service by enabling the AI model to automatically label 3D objects without continuous human intervention. Once trained on a small labelled subset, the model independently processes and labels the entire dataset, making the system self-sufficient for large-scale data generation and eliminating the need for manual labelling of every object.
3Quantity of substance
If one single labelled 3D object is segmented into many 2D images from different angles, then many training images can be obtained, but the 3D object still needs to be manually labelled first
Solution Approach 1:
The patent performs preliminary action by manually labelling only a small subset of 3D objects first, then uses this initial labelled data to train an AI model that automatically labels the remaining objects. This preliminary manual labelling provides the foundation for subsequent automated labelling, dramatically reducing total labelling time while maintaining accuracy.
Solution Approach 2:
The patent introduces an AI model as an intermediary between manual labelling and final labelled dataset creation. The AI model is trained on manually labelled data and then used to automatically label remaining objects, serving as a mediator that transfers labelling knowledge from human experts to the entire dataset without requiring direct human involvement in labelling every object.
Data Source
AI summary
There is provided a method for generating labelled 3D data representations of real-world objects, comprising: —obtaining, using processing circuitry (110), a set of one or more pre-labelled 3D coordinate, each 3D coordinate in the set of one or more pre-labelled 3D coordinate representing a point on a real world object in a real world area, —for each of the one or more pre-labelled 3D coordinates in the obtained set: —obtaining a template volume to be applied to the 3D coordinate using the processing circuitry (110); —generating an object volume comprising the 3D coordinate, using the processing circuitry (110), by applying the obtained template volume to the 3D coordinate; and—labelling the generated object volume, by assigning the object label of the pre-labelled 3D coordinate to all 3D coordinates located within the generated object volume. Also provided are a system and computer program product.