Object Recognition Learning Data from Geometric Imaging Views

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for generating learning data for object recognition require significant effort and precise model data, especially when increasing the variation of photo-presentations, and often necessitate actual image processing or computer graphics-based virtual image creation, which can be cumbersome and require precise adjustments to match real-world imaging environments.

Innovation Solution

An information processing device that acquires geometric information about a target object, decides optimal imaging positions and orientations, and generates learning data using this information and acquired images, allowing for efficient creation of adaptable learning data for object recognition applications.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If image processing methods (enlargement/reduction, rotation) are applied to increase variation of photo-presentations, then the quantity of learning data is improved, but the complexity of manual work and time consumption increase

Engineering Contradiction:

Improvequantity of learning dataVSAvoidtime consumption for generating learning data

Core Design Contradiction:

Quantity of substanceVSLoss of time

Solution Approach 1:

The patent creates virtual copies of the target object through 3D modeling and renders multiple views automatically. Instead of manually processing each image variation, the system generates synthetic training images by rendering the 3D model from different angles, distances, and lighting conditions, dramatically reducing manual time investment while producing abundant learning data

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent performs preliminary 3D scanning and modeling of the target object before generating learning data. By creating a complete 3D digital model in advance, the system enables automatic generation of numerous view variations without repeated manual imaging, allowing rapid production of diverse training images from a single preprocessed model

Inventive Principle:

Principle #10Preliminary action

2Quantity of substance

If CG-based virtual image creation is used to increase variation, then the quantity of learning data is improved, but the precision of model data and complexity of environment adjustment increase

Engineering Contradiction:

Improvequantity of learning dataVSAvoidcomplexity of model data and environment adjustment

Core Design Contradiction:

Quantity of substanceVSDevice complexity

Solution Approach 1:

The system creates accurate 3D digital copies of the actual target object through scanning technology. This precise digital replica captures the true geometry and characteristics of the object, eliminating the need for manual model construction and ensuring high fidelity without requiring complex environment adjustments, as the 3D model itself contains all necessary geometric information

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent replaces complex mechanical CG environment setup with automated rendering processes. Instead of manually adjusting virtual lighting, cameras, and environments in CGI software, the system uses programmatic rendering of the 3D model with automatically generated view parameters, significantly reducing the complexity of environment adjustment while maintaining data quality

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Measurement precision

If actual imaging from various positions and orientations is performed to generate learning data, then the accuracy of recognition is improved, but the workload and cost increase

Engineering Contradiction:

Improveaccuracy of recognitionVSAvoidworkload and cost of data collection

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The patent uses 3D scanning to create a digital copy of the target object, which can then be virtually viewed from any position and orientation without physical relocation. This digital copying approach maintains the accuracy benefits of multi-angle imaging while eliminating the need for physical repositioning, dramatically reducing workload and costs associated with collecting images from various perspectives

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The system enables the 3D model to serve itself by automatically generating all necessary view variations through computational rendering. The single scanned model contains all geometric information needed to produce unlimited view variations, making the system self-sufficient for generating diverse training data without requiring additional physical imaging resources or manual intervention

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS10997465B2Information processing device, information processing method, and storage medium

Publication Date: 2021.05.04 CANON KK

AI summary

An information processing device includes a first acquiring unit configured to acquire geometric information relating to a target object to be recognized, a decision unit configured to decide an imaging position and orientation at which the target object is imaged, based on the geometric information acquired by the first acquiring unit, a second acquiring unit configured to acquire an image of the target object which has been captured at the imaging position and orientation decided by the decision unit, and a generation unit configured to generate learning data, based on the geometric information acquired by the first acquiring unit and the image acquired by the second acquiring unit.