3D Training Sample Generation for Semantic Segmentation Labeling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The manual preparation of training samples for deep learning models to recognize character images in games is inefficient and costly, requiring significant time and effort for manual screenshots and labeling.

Innovation Solution

A method and apparatus for automatically generating training samples for semantic segmentation models by using a three-dimensional model, rotating it to different angles, applying colorful skin pictures, and creating label images to train the model, eliminating the need for manual screenshots and labeling.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If manual screenshots and labeling are used to prepare training samples, then the training data can be obtained, but the process is inefficient and costly, requiring significant time and effort

Engineering Contradiction:
Improvetraining sample generation efficiencyVSAvoidtime required for manual preparation
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent uses a three-dimensional model as a copy or representation of the actual game character, which can be rendered from multiple angles and perspectives. This digital model replaces the need for manual screenshot capture and labeling, as the model can automatically generate training images with precise segmentation masks through rendering, thereby dramatically improving efficiency and reducing time consumption.

Inventive Principle:
Principle #26Copying

2Manufacturing precision

If manual labeling is performed for each screenshot, then accurate training data with labels can be obtained, but the process is time-consuming and labor-intensive

Engineering Contradiction:
Improvelabeling accuracyVSAvoidsample generation speed
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The patent replaces the manual mechanical process of screenshot capture and labeling with an automated rendering system. The three-dimensional model is rendered from various angles and perspectives to generate training images, and the segmentation masks are automatically created based on the model's geometric data and material properties, eliminating the need for manual labeling while maintaining high accuracy and improving productivity.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Adaptability or versatility

If diverse training samples are collected manually from game scenes, then the model can learn various character appearances, but the process requires significant human effort and time

Engineering Contradiction:
Improvecharacter recognition coverageVSAvoidtime for sample collection
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent employs a dynamic rendering approach where the three-dimensional model can be rotated and viewed from multiple angles, and different lighting conditions and backgrounds can be applied. This allows the generation of diverse training samples covering various character appearances, poses, and environments automatically, improving adaptability without requiring manual collection of diverse scenes.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3989120B1Method and device for generating training sample for semantic segmentation model, storage medium, and electronic device
Publication Date: 2026.04.15 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3989120B1 patent drawingFigure 1~2A
  • EP3989120B1 patent drawingFigure 2B~2C
  • EP3989120B1 patent drawingFigure 2D~2E

AI summary

A method and device for generating a training sample for a semantic segmentation model, a storage medium, and an electronic device. The semantic segmentation model is used for segmenting an object in an image. The method for generating a training sample for the semantic segmentation model is executed by the electronic device and comprises: acquiring an object identifier of an object to be trained and a model file corresponding to the object identifier, the model file comprising a three-dimensional model and a skin texture set of said object, and the skin texture set comprising multiple pure color skin textures and multiple color skin textures (S101); determining transformation models of the three-dimensional model in different orientation angles, each of the orientation angles corresponding to one of the transformation models (S 102); generating, on the basis of the multiple color skin textures and each transformation model, multiple sample image sets corresponding to each transformation model (S 103); generating annotation images of each of the sample image sets on the basis of the object identifier, of the multiple pure color skin textures, and of each transformation model (S 104); and generating a training sample of said object on the basis of the annotation images and of the sample image sets (S 105), thus training the semantic segmentation model via the training sample of said object.