Synthetic Image Training Dataset Generation via 3D Rendering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional data labeling for machine learning requires manual operation, leading to time-consuming and inaccurate labeling of large numbers of photos, resulting in deviations in labeling accuracy.
Innovation Solution
A system and method that utilize three-dimensional data to generate training datasets by creating synthetic images and extracting labeling information, employing deep learning techniques with encoders, style transfer modules, and decoders to automate the labeling process, reducing time and ensuring uniformity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual labeling operation is used, then labeling accuracy can be maintained, but labeling time becomes extremely long and productivity is low
Solution Approach 1:
The patent creates synthetic images that copy and replicate real-world images with labeled targets. These synthetic images are generated by rendering 3D models with various parameters (lighting, camera angles, backgrounds) to produce training data that mirrors real imaging conditions, thereby eliminating the need for time-consuming manual labeling of numerous real photos
Solution Approach 2:
The patent replaces the mechanical manual labeling process with an automated computer-based system. A controller automatically generates synthetic images from 3D data and extracts labeling information through computational algorithms, substituting human operators with automated processing mechanisms that operate continuously without fatigue or variability
2Measurement precision
If manual labeling is performed by operators, then labeling can be done with current technology, but labeling accuracy deviates depending on operator performance and consistency is poor
Solution Approach 1:
The system performs self-service by automatically generating its own training data without human intervention. The controller autonomously renders synthetic images from 3D models and extracts labeling information through automated algorithms, ensuring that the labeling process is independent of human performance variations and maintains consistent accuracy across all training samples
Solution Approach 2:
The patent changes the parameters of image generation by using adjustable 3D model parameters (geometry, material properties, lighting conditions, camera parameters) to create diverse synthetic images. This allows systematic control over the labeling accuracy and consistency by optimizing the rendering parameters and extraction algorithms rather than relying on operator skill levels
3Quantity of substance
If large number of photos are used for training data, then machine learning performance improves, but the time and resources required for labeling increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-generating a large library of synthetic training images before the actual machine learning training begins. The system renders numerous synthetic images with various parameters and extracts labeling information in advance, creating a ready-to-use training dataset that can be immediately utilized without time-consuming labeling operations during the training process
Data Source
AI summary
Provided is a training dataset generating system including: a communicator to receive a two-dimensional (2D) image obtained by photographing a target object; and a controller configured to generate, based on the 2D image and based on three-dimensional (3D) data for the target object, a training dataset comprising a synthetic image and comprising labeling information, wherein the controller is configured to generate the training data set by: generating, based on the 3D data, a rendered image, generating the synthetic image, based on the 2D image and the rendered image, through deep learning training, extracting, based on at least one of the 3D data or the rendered image, the labeling information for the target object, and generating the training dataset.