Training Data for Visual Quality Inspection

It is the labeling of production images to show a model which objects, regions, or conditions matter. Unitlab supports defect localization, surface segmentation, and image-level condition labels for training and evaluation datasets.

Yes. Define image properties for the inspection outcome, then add geometry where a visible defect is present. A clear rule for defect-free examples helps reviewers apply the same standard.

Teams can define classes for visible scratches, cracks, dents, missing components, assembly issues, and packaging damage. The labeling scheme should reflect the actual inspection objective and what can be observed in the images.

Use boxes when the task needs a defect location. Use polygons or masks when the model needs the precise damaged area. The intended model output and the defect shape determine the right label type.

This page focuses on defects and conditions in individual inspection images. Process-video datasets label how components, machine states, or assembly steps change across frames.

Define shared classes, structured properties, and clear labeling instructions before work starts. Use representative examples and contextual review to resolve disagreements in the industrial visual inspection dataset.

Yes. Route image annotation through Review and Rework stages. Reviewers can inspect the source data, correct labels, and send an item back when more work is needed.

Yes. Group related inspection images, process video, supported sensor recordings, or notes into a shared item when the task needs that context. Review each file in its compatible viewer and use shared definitions for the inspection finding.

Export reviewed image annotations in a supported format appropriate to the label types. Dataset versions help teams identify which prepared examples belong to the training or evaluation release.