Domain Adaptation via Pixel-Aligned Synthetic Data

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

The high cost and labor intensity of generating human-labeled training data for neural networks, along with the domain gap issues when using synthetic data, lead to poor performance in real-world applications.

Innovation Solution

A system that generates synthetic training images with pixel-level alignment using a calibration object with fiducial markers, allowing for effective training of image processing neural networks to bridge the domain gap between synthetic and real data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If synthetic training data is used to train neural networks, then the cost and time of data collection are reduced, but domain gap between synthetic and real images causes poor performance in real-world applications

Engineering Contradiction:

Improvetime required for manual labeling processVSAvoidperformance of machine learning model on real data

Core Design Contradiction:

Loss of timeVSReliability

Solution Approach 1:

The patent introduces a calibration object with fiducial markers as an intermediary element that appears in both real and synthetic images. This mediator enables precise alignment and domain adaptation by providing common reference points across domains, allowing the model to learn domain-invariant features while training primarily on synthetic data

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent transforms synthetic images to match the statistical properties and characteristics of real images by adjusting parameters such as noise levels, lighting conditions, and distortion patterns based on analysis of real image data. This parameter transformation reduces the domain gap while maintaining the efficiency of synthetic data generation

Inventive Principle:

Principle #35Parameter changes

2Measurement precision

If human reviewers are used to label training data, then accurate labels are obtained, but the process is expensive and time-consuming

Engineering Contradiction:

Improveaccuracy of labelsVSAvoidcost and time of labeling process

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The patent creates synthetic copies of real images with automatically generated labels that replicate the labeling task. By generating synthetic images with known ground truth labels through computer graphics rendering, the system obtains accurate labels without human intervention, dramatically improving productivity while maintaining label accuracy

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The system uses automated computer vision algorithms and computer graphics rendering to self-generate labeled training data without human reviewers. The synthetic data generation process automatically produces images with precise ground truth annotations, making the labeling process self-service and eliminating dependency on human annotators

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS11170581B1Supervised domain adaptation

Publication Date: 2021.11.09 INTRINSIC INNOVATION LLC

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a feature extraction neural network to generate domain-invariant feature representations from domain-varying input images. In one aspect, the method includes obtaining a training dataset comprising a first set of target domain images and a second set of real domain images that each have pixel-wise level alignment with a corresponding target domain image, and training the feature extraction neural network on the training dataset based on optimizing an objective function that includes a term that depends on a similarity between respective feature representations generated by the network for a pair of target and source domain images.