P
pigzippa47
Guest
- Thread Author
- #1
This blog post is reprinted here with the permission of Yazaki Innovation.
As computer vision systems become more sophisticated, the quality of their training data becomes increasingly important. Yet annotation quality is not simply a matter of whether a bounding box, polygon, or segmentation mask is technically correct. It is also a process-management challenge.
Different annotators may interpret the same object differently. Ambiguous requirements can create rework, while complex images involving occlusion, poor lighting, unclear boundaries, or rare edge cases can increase variability.
At Yazaki, we approach this challenge using principles developed through more than 80 years of manufacturing experience. In labor-intensive manufacturing, consistent quality depends on standardized processes, trained people, operational discipline, knowledge retention, and systematic inspection. Many of these same principles can be applied to producing reliable AI training data.
Before production begins, requirements need to be translated into clear annotation rules. What constitutes an object? Where should its boundary begin and end? How should partially occluded objects be handled? What happens when an image contains a situation that was not anticipated in the original specification?
Yazaki addresses these questions through requirement alignment and trial stages before full-scale annotation begins. Detailed class definitions and annotation guidelines are established, while undefined or ambiguous cases are identified through Q&A and incorporated into the working rules.
The objective is to reduce interpretation differences before they become thousands of inconsistent labels.
Yazaki uses trained employees for its annotation operations rather than relying on a distributed crowdsourcing model. Employees assigned to annotation work undergo Yazaki’s training and qualification process, while supervisors provide guidance during execution.
Standardized procedures also help retain project knowledge within the organization as teams and projects scale.
This reflects a principle long understood in manufacturing: a repeatable process requires both standardized procedures and people trained to execute them consistently.
For projects requiring Yazaki’s highest level of quality control, the workflow can incorporate a three-stage inspection structure. Following annotation, a full inspection is conducted at the operational level. Additional quality checks are then performed at higher levels of the project organization, with the final stage able to include review by the Japan-based project team.
The number and type of inspections can also be adjusted according to project requirements, complexity, and budget.
This multi-layer approach is designed to identify not only individual annotation errors, but also systematic differences in how requirements are being interpreted across the workforce.
A single road scene can contain vehicles, lane markings, road boundaries, signs, infrastructure, and other classes simultaneously. Lighting variation and occlusion can make these elements difficult to label consistently across a large dataset.
In one autonomous-driving AI application, Yazaki performed semantic segmentation of complex road scenes. Detailed class definitions and quality-control processes were established to maintain consistent labeling of multiple classes, including lane markings and road boundaries under varying visual conditions.
The domain changes, but the underlying challenge remains similar: humans must convert complex visual information into structured training data according to consistent rules.
The same principles that help manufacturers produce physical components consistently—standardization, workforce training, disciplined execution, inspection, and continuous feedback—can also help computer vision teams build more reliable training datasets.
https://service.yazaki.ai/annotation
The post How Manufacturing Quality-Control Practices Can Improve Computer Vision Training Data appeared first on Edge AI and Vision Alliance.
As computer vision systems become more sophisticated, the quality of their training data becomes increasingly important. Yet annotation quality is not simply a matter of whether a bounding box, polygon, or segmentation mask is technically correct. It is also a process-management challenge.
Different annotators may interpret the same object differently. Ambiguous requirements can create rework, while complex images involving occlusion, poor lighting, unclear boundaries, or rare edge cases can increase variability.
At Yazaki, we approach this challenge using principles developed through more than 80 years of manufacturing experience. In labor-intensive manufacturing, consistent quality depends on standardized processes, trained people, operational discipline, knowledge retention, and systematic inspection. Many of these same principles can be applied to producing reliable AI training data.
1. Quality Starts Before the First Annotation
One common source of annotation inconsistency is not the annotator. It is ambiguity in the requirements.Before production begins, requirements need to be translated into clear annotation rules. What constitutes an object? Where should its boundary begin and end? How should partially occluded objects be handled? What happens when an image contains a situation that was not anticipated in the original specification?
Yazaki addresses these questions through requirement alignment and trial stages before full-scale annotation begins. Detailed class definitions and annotation guidelines are established, while undefined or ambiguous cases are identified through Q&A and incorporated into the working rules.
The objective is to reduce interpretation differences before they become thousands of inconsistent labels.
2. Train the Workforce, Not Just the AI Model
Consistency also depends on who performs the annotation.Yazaki uses trained employees for its annotation operations rather than relying on a distributed crowdsourcing model. Employees assigned to annotation work undergo Yazaki’s training and qualification process, while supervisors provide guidance during execution.
Standardized procedures also help retain project knowledge within the organization as teams and projects scale.
This reflects a principle long understood in manufacturing: a repeatable process requires both standardized procedures and people trained to execute them consistently.
3. Build Inspection Into the Production Process
Quality control should not be limited to reviewing a small sample at the end of a project.For projects requiring Yazaki’s highest level of quality control, the workflow can incorporate a three-stage inspection structure. Following annotation, a full inspection is conducted at the operational level. Additional quality checks are then performed at higher levels of the project organization, with the final stage able to include review by the Japan-based project team.
The number and type of inspections can also be adjusted according to project requirements, complexity, and budget.
This multi-layer approach is designed to identify not only individual annotation errors, but also systematic differences in how requirements are being interpreted across the workforce.
Figure 1 — Yazaki’s three-stage annotation quality-control process: full inspection followed by additional management-level quality checks.
Across high-complexity annotation projects, Yazaki has recorded more than 99.7% annotation accuracy, including work involving low-resolution imagery, ambiguous boundaries, microscopic images, complex satellite imagery, and rare edge cases.4. Applying the Process to Complex Computer Vision Data
Autonomous-driving training data provides a useful example of why process discipline matters.A single road scene can contain vehicles, lane markings, road boundaries, signs, infrastructure, and other classes simultaneously. Lighting variation and occlusion can make these elements difficult to label consistently across a large dataset.
In one autonomous-driving AI application, Yazaki performed semantic segmentation of complex road scenes. Detailed class definitions and quality-control processes were established to maintain consistent labeling of multiple classes, including lane markings and road boundaries under varying visual conditions.
Figure 2 — Example of semantic segmentation for autonomous-driving AI: original road scene and annotated training data.
The same process principles can be applied to other complex visual datasets. Yazaki has annotation experience across geospatial imagery, factory environments, agricultural imagery, pathology images, and engineering drawings.The domain changes, but the underlying challenge remains similar: humans must convert complex visual information into structured training data according to consistent rules.
5. What Should AI Teams Look for in an Annotation Process?
When evaluating an annotation workflow or provider, AI teams should look beyond cost per image or annotation speed. Four questions are particularly important:- How are ambiguous requirements converted into standardized annotation rules?
- How are annotators trained and qualified before working on production data?
- How are interpretation differences identified and incorporated back into the guidelines?
- Does quality control rely primarily on sampling, or are additional inspection layers available for complex or high-accuracy datasets?
The same principles that help manufacturers produce physical components consistently—standardization, workforce training, disciplined execution, inspection, and continuous feedback—can also help computer vision teams build more reliable training datasets.
Learn more about Yazaki Image Annotation Services:
https://service.yazaki.ai/annotation
The post How Manufacturing Quality-Control Practices Can Improve Computer Vision Training Data appeared first on Edge AI and Vision Alliance.