2D Feature Database Generation from 3D Content
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional learning-based approaches for AI engines face challenges in generating accurate and efficient training data due to the scarcity and high cost of large image or video databases, requiring significant human effort for annotation and struggling with scene characteristics difficult to capture, such as object boundary and position prediction.
Innovation Solution
A method and system that acquire 3D content, project it to 2D space, and determine latent variables to generate a 2D feature database, utilizing synthetic 3D objects and processing algorithms to create ground truth data for training AI engines, reducing the need for extensive human annotation and improving accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If large amounts of image or video content are used for training data, then the AI engine can be trained, but the cost increases and accuracy may be compromised
Solution Approach 1:
The patent creates synthetic training data by projecting 3D objects into 2D images, generating realistic training samples without requiring large amounts of real image or video content. This copying approach from 3D models to 2D images provides sufficient training data while avoiding the costs and quality issues of collecting and processing large volumes of real-world images
Solution Approach 2:
The patent transitions from 3D object space to 2D image space by projecting three-dimensional objects onto two-dimensional planes. This dimensional transformation allows the system to generate training data in the required 2D format while leveraging the structured information available in 3D space, thereby improving both efficiency and accuracy
2Ease of manufacture
If human labor is used for annotating ground truth data, then training data can be created, but significant effort and resources are required
Solution Approach 1:
The system automatically generates ground truth data by computing 2D projections of 3D objects and extracting relevant features through algorithmic processing. This self-service approach eliminates the need for manual human annotation, simultaneously reducing effort requirements and increasing productivity through automated batch processing
Solution Approach 2:
The patent replaces the mechanical process of manual human annotation with automated computational algorithms that project 3D objects to 2D images and extract features programmatically. This substitution eliminates human labor while maintaining or improving the quality and speed of training data generation
3Adaptability or versatility
If image or video content with scene characteristics is used, then real-world scenarios can be captured, but these characteristics are difficult to annotate or label
Solution Approach 1:
Instead of attempting to annotate complex scene characteristics in real images, the patent inverts the approach by generating images from known 3D objects with defined properties. This allows the system to control and know the ground truth characteristics a priori, making annotation trivial since the ground truth is embedded in the 3D object definitions rather than extracted from complex scenes
Data Source
AI summary
One embodiment provides a method comprising acquiring 3D content comprising a 3D object in 3D space. The 3D object has object information indicative of a location of the 3D object in the 3D space. The method further comprises projecting the 3D object to a 2D object in 2D space based on the object information. The 2D object has one or more 2D vertices indicative of a location of the 2D object in the 2D space. The method further comprises determining one or more latent variables in the 2D space based on the object information and the one or more 2D vertices, and generating a 2D feature database including the one or more latent variables.