Target Detection Model Updating with 3D Synthetic Training Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In intelligent warehousing automation, existing target detection models for picking robots lack efficient methods for updating their detection precision, especially when dealing with varying angles and scenarios.
Innovation Solution
A method and apparatus for updating a target detection model by constructing a three-dimensional model of a target item from image data at multiple angles, generating a composite image, and using this image as a sample to train the model, thereby improving detection precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional target detection models are used for picking robots, then the detection process is simple, but the detection precision is insufficient when dealing with varying angles and scenarios
Solution Approach 1:
The patent transforms two-dimensional image data into three-dimensional target models, adding a spatial dimension to the detection process. This allows the model to understand target geometry, orientation, and spatial relationships, significantly improving detection precision for objects at varying angles while maintaining manageable complexity through automated 3D reconstruction algorithms.
Solution Approach 2:
The patent creates synthetic training samples by generating composite images that combine target models with background images. These copied and synthesized images provide diverse training data without requiring physical photographs of every possible scenario, improving detection precision while avoiding the complexity of extensive data collection.
2Measurement precision
If more training data is collected to improve detection precision, then the model performance improves, but the time and resources required for data collection and model updating increase
Solution Approach 1:
The system performs self-updating by automatically generating its own training data. The target detection model uses detected targets to generate 3D models, which then produce synthetic training images. This self-service mechanism eliminates the need for manual data collection and annotation, improving detection precision without increasing model updating time.
Solution Approach 2:
The patent pre-generates three-dimensional target models and composite images that can be used as training data before actual detection tasks. This preliminary preparation of diverse training samples ensures the model is ready for various scenarios, improving detection precision while avoiding time-consuming data collection during operation.
3Measurement precision
If the target detection model is updated frequently to maintain high precision, then the detection accuracy improves, but the computational resources and complexity of the system increase
Solution Approach 1:
The system creates synthetic copies of target images through 3D model rendering and composite image generation. These copied images serve as training data, allowing frequent model updates with high detection accuracy while avoiding the complexity of collecting and processing large amounts of real-world image data.
Solution Approach 2:
By transitioning from 2D images to 3D models and back to synthesized 2D images, the system creates diverse training samples from limited original data. This dimensional transformation enables frequent model updates with high accuracy while keeping the underlying 3D model representation computationally efficient and manageable.
Data Source
AI summary
Disclosed in the present disclosure are a method and apparatus for updating a target detection model. A specific implementation of the method comprises: constructing a three-dimensional model of a target article according to image data of the target article at a plurality of angles; generating, according to the three-dimensional model, a synthetic image comprising a target article object, which represents the target article; by taking the synthetic image as a sample image and taking the target article object as a label, obtaining training samples to generate a training sample set; and training a target detection model by means of the training sample set, so as to obtain an updated target detection model.