Synthetic Training Image Generation for Visual Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current visual object recognition systems face challenges in accurately identifying objects from images due to background clutter and varying lighting conditions, which can lead to reduced matching accuracy and efficiency.
Innovation Solution
The method involves generating synthetic training images based on 3D object data models from multiple virtual views, allowing for clean and clutter-free images that can be matched using a visual object recognition module, which associates information about the object and virtual view with the matched images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If real images with background clutter and varying lighting conditions are used for training, then the visual object recognition system can handle real-world scenarios, but the matching accuracy decreases due to background clutter and lighting variations
Solution Approach 1:
The patent creates synthetic training images by rendering 3D object models with controlled lighting and without background clutter. These synthetic copies serve as clean training data that preserves object characteristics while eliminating harmful factors like background interference and inconsistent lighting, thereby improving matching accuracy without sacrificing real-world applicability
Solution Approach 2:
The patent introduces 3D object models as an intermediary between real objects and training images. By scanning real objects to create 3D models and then rendering synthetic images from these models, the system obtains clean training data that captures essential object features while removing background clutter and lighting variations that hinder accurate matching
2Measurement precision
If synthetic training images are generated from 3D object data models, then matching accuracy improves by eliminating background clutter and lighting variations, but the complexity of the system increases due to 3D modeling and rendering requirements
Solution Approach 1:
The patent performs preliminary actions by creating 3D object data models in advance through scanning real objects. These pre-created 3D models serve as reusable sources for generating synthetic training images, eliminating the need to process real images with clutter and variable lighting during the actual recognition task, thereby improving accuracy while managing complexity through upfront preparation
3Measurement precision
If multiple virtual views are used to render synthetic training images, then the semantic understanding and tagging accuracy improves, but the rendering time and computational resources increase
Solution Approach 1:
The patent renders synthetic training images from multiple virtual views around the 3D object model, capturing the object from different angles and perspectives. This partial rendering approach (focusing on key viewpoints) provides sufficient semantic information for accurate understanding and tagging while avoiding the need to render all possible views, thereby balancing accuracy improvement with reasonable rendering time
Data Source
AI summary
Methods and systems for rendering virtual views of three-dimensional (3D) object data models are described. An example method may include receiving information associated with rendering a 3D object data model, and determining virtual views for rendering the 3D object data model. Based on the virtual views and the received information, synthetic training images of the object may be rendered, and information identifying a given virtual view used to render a synthetic training image may be stored with the synthetic training images. A visual object recognition module may determine from a plurality of images and image that substantially matches one or more of the synthetic training images, and the matched image may be associated with information identifying the object and a virtual view used to render a matching synthetic training image. Another method may also leverage information from a synthetic training image to help recognize other objects of the matched image.