3D Model Image Augmentation for ML Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training machine learning models, especially image-based models, face challenges in acquiring effective training data, as finding relevant images can be difficult and time-consuming, particularly for specific objects like steel pipes with defects.
Innovation Solution
A system that generates a three-dimensional (3D) model of an object from images using machine learning, captures snapshots at different angles, fuses features into these snapshots, and stores them for use as training data, leveraging cloud computing and image-based search engines to automate the process of acquiring large volumes of training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a data scientist manually searches for training images of objects with defects, then the training data can be obtained, but the time and effort required becomes excessive
Solution Approach 1:
The system pre-generates a comprehensive 3D model of the object containing all possible defect variations before actual training data is needed. This preliminary action creates a ready-to-use defect library that can be quickly accessed and rendered into training images without manual searching, thus resolving the contradiction between data quality and acquisition time
Solution Approach 2:
Instead of manually collecting real defect images, the system creates synthetic copies of defects by rendering them onto 3D model surfaces from various angles and conditions. These synthesized training images provide diverse defect examples without requiring extensive manual image collection, thereby reducing time while maintaining training data quality
2Adaptability or versatility
If many different images with various defects are collected for training, then the machine learning model can be trained effectively, but finding such images becomes increasingly difficult
Solution Approach 1:
The system transitions from 2D image collection to 3D model-based synthesis. By creating a 3D model and rendering images from numerous virtual camera angles, lighting conditions, and defect positions, the system generates diverse training data with varying defect types and orientations, achieving high adaptability while simplifying the data acquisition process
Solution Approach 2:
The system dynamically generates training images by programmatically varying parameters such as camera angles, lighting conditions, defect positions, and defect types. This dynamic generation process creates diverse training data without manual collection, making it easy to produce adaptable training sets for different defect scenarios
3Quantity of substance
If a large corpus of training data is prepared, then the machine learning model training can proceed, but the process of collecting and preparing this data becomes time-consuming
Solution Approach 1:
The system efficiently produces large volumes of training data by synthesizing images from a single 3D model through virtual rendering. By copying the defect features onto different views and conditions of the 3D model, the system generates extensive training datasets without the time-consuming process of manually collecting and preparing individual images
Solution Approach 2:
The system performs preliminary processing by creating the 3D model and defect library once, then rapidly generates additional training images on-demand through rendering. This preliminary action eliminates the need for repeated manual data collection, significantly improving the productivity of training data preparation while maintaining large data volumes
Data Source
AI summary
An example operation may include one or more of generating a three-dimensional (3D) model of an object via execution of a machine learning model on one or more images of the object, capturing a plurality of snapshots of the 3D model of the object at different angles to generate a plurality of snapshot images of the object, fusing a feature into each of the plurality of snapshots to generate a plurality of fused snapshots of the 3D model of the object, and storing the plurality of fused snapshots of the 3D model of the object in memory.