3D Model Rendering Parameter Changes for Teacher Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for generating teacher data for deep learning in image recognition face challenges such as high costs due to the need for numerous 3D models, limited variations in texture and shape, and inability to create data with partially masked recognition targets, leading to poor recognition accuracy.
Innovation Solution
An image processing method that dynamically changes rendering parameters in 3D models to generate diverse teacher data, including variations in texture, shape, camera angles, and masking ratios, thereby increasing the quantity and variability of teacher data without redundancy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional 3D model methods are used to generate teacher data, then the generation process is simple, but the variations in texture and shape are limited and recognition accuracy is poor
Solution Approach 1:
The patent applies parameter changes by dynamically modifying rendering parameters including texture parameters, shape parameters, camera parameters, and lighting parameters. This allows the system to generate diverse teacher data with varied textures and shapes from a single 3D model, resolving the contradiction between generation simplicity and data diversity.
Solution Approach 2:
The system introduces dynamics by randomly changing rendering parameters during the data generation process. Instead of static 3D models with fixed appearances, the system dynamically generates multiple variations by adjusting parameters such as texture coordinates, shape deformations, camera angles, and lighting conditions, thereby achieving both simplicity and versatility.
2Adaptability or versatility
If multiple 3D models are collected to increase variations, then more teacher data can be generated, but costs increase significantly
Solution Approach 1:
The patent applies copying by generating multiple variations of teacher data from a single 3D model through parameter changes. Instead of collecting multiple expensive 3D models, the system creates numerous copies with different textures, shapes, and rendering conditions, thereby reducing the quantity of original models needed while maintaining high versatility.
Solution Approach 2:
By systematically changing rendering parameters such as texture maps, shape coefficients, camera positions, and lighting setups, the system generates diverse teacher data from one base model. This parameter-based approach replaces the need for collecting multiple 3D models, significantly reducing costs while preserving data variety.
3Adaptability or versatility
If data augmentation through image processing is used, then variations can be increased, but images of side or rear views cannot be generated from front views
Solution Approach 1:
The patent applies dynamics by enabling dynamic camera parameter changes within the 3D rendering system. Instead of being limited to fixed front-view images that require complex post-processing for augmentation, the system dynamically adjusts camera angles, positions, and orientations to generate authentic multi-angle views including side and rear perspectives, thereby improving both versatility and operational ease.
Solution Approach 2:
The system transitions from two-dimensional image processing to three-dimensional rendering by changing the dimensionality of the data generation approach. By working in 3D space with adjustable camera parameters, the system can naturally generate views from any angle including side and rear perspectives, overcoming the limitations of 2D data augmentation methods.
4Reliability
If teacher data with partially masked recognition targets is required, then recognition robustness improves, but conventional 3D model methods cannot generate such data
Solution Approach 1:
The patent applies segmentation by introducing masking parameters that can selectively obscure different portions of the recognition target in generated teacher data. This allows the system to create partially masked images with controlled visibility of specific features, enabling robustness training while maintaining the ability to generate diverse masking patterns through parameter adjustments.
Solution Approach 2:
By adding masking parameters to the rendering parameter set, the system can dynamically control which portions of the recognition target are visible or obscured in generated images. This parameter-based approach enables flexible generation of partially masked targets with varying degrees and patterns of occlusion, improving recognition robustness without limiting versatility.
Data Source
AI summary
An image processing method performed by a computer, the method including: changing a rendering parameter in a three-dimensional model of a recognition target; and generating teacher data of the recognition target based on the rendering parameter changed by the changing. Thereby, sufficient amount of teacher data can be obtained, without any unfavorable redundancy.