Neural Network Object Detection Feature Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection technologies struggle to accurately recognize objects in images using full image information, particularly when objects are occluded or have similar shapes.
Innovation Solution
A method that involves obtaining an image and extracting class, pose, and relationship features of objects within the image. These features are then corrected using a combination of weights and shared across sub-networks within a neural network, allowing for improved object detection and virtual object generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If object detection uses full image information, then detection accuracy is improved, but processing complexity increases
Solution Approach 1:
The patent divides the object detection task into multiple sub-networks, each responsible for extracting specific features (class features, pose features, relationship features). This segmentation allows the system to process full image information systematically while managing computational complexity through specialized processing units.
Solution Approach 2:
The patent transforms 2D image data into multiple feature dimensions by extracting class, pose, and relationship features from intermediate layers of sub-networks. This dimensional transformation enables the system to capture comprehensive object information while organizing processing through structured feature representations.
2Measurement precision
If the system extracts and corrects multiple feature types, then object information accuracy is improved, but computational time increases
Solution Approach 1:
The patent extracts features from intermediate layers of sub-networks before final classification, performing preliminary feature extraction and correction. This allows the system to prepare and refine object information in advance, improving accuracy while optimizing the timing of computational operations.
Solution Approach 2:
The patent implements a correction mechanism where features are refined using feedback from multiple sub-networks. The class features, pose features, and relationship features are corrected based on interactions between sub-networks, allowing iterative improvement of object information accuracy.
3Manufacturing precision
If virtual objects are generated with multiple property sets, then realism is improved, but user selection complexity increases
Solution Approach 1:
The patent generates multiple sets of virtual object properties (position, pose, action) that may exceed what is strictly necessary. This excessive generation of candidate properties allows the system to provide comprehensive virtual object options, from which users can select the most appropriate properties for their specific needs.
Data Source
AI summary
An electronic device for estimating object information and generating a virtual object and a method of operating the electronic device are disclosed. The method includes obtaining an image, obtaining a class feature, a pose feature, and a relationship feature of an object included in the image, correcting each of the class feature, the pose feature, and the relationship feature using any combination of any two or more of the class feature, the pose feature, and the relationship feature of the object, and obtaining class information, pose information, and relationship information of the object based on the corrected class feature, the corrected pose feature, and the corrected relationship feature, respectively.