# How do you test egg vision with synthetic data? > Synthetic data with ground truth are egg scenes generated by computer, in which the true class of every egg is already known. On those scenes Vigiovo got 98.7% of tracked eggs right on video (376 of 381) and about 95% on a single frame, and rejected no clean eggs. These numbers measure the system, not your line: production accuracy depends on your camera, your light and your eggs. Updated: 2026-10-09 · URL: https://vigiovo.com/en/resources/synthetic-ground-truth-data · Português: https://vigiovo.com/recursos/dados-sinteticos-com-gabarito.md ## Key takeaways - Synthetic scenes have ground truth: the true class of each egg is known, so accuracy can be counted egg by egg. - On tracked video, Vigiovo got 98.7% of eggs right (376 of 381) and rejected no clean eggs in the synthetic tests. - On a single frame, accuracy per egg was about 95%. - Synthetic accuracy is not production accuracy: your line needs [calibration with labelled eggs](https://vigiovo.com/en/resources/calibrating-with-labelled-eggs). ## What is synthetic data with ground truth? Synthetic data are images generated by computer, not photos of a conveyor. What makes them useful for testing is the ground truth: the scene is built with known defects, and the class of each egg (clean, dirty, cracked) is already recorded. A real photo does not carry that answer, so someone has to label it by hand, with all the human error that brings. With ground truth, the comparison can be automatic. The system states the class of each egg, the test compares that answer with the recorded one and counts the matches. The process is repeatable: the same scene with the same configuration gives the same result, because Vigiovo uses [deterministic computer vision](https://vigiovo.com/en/resources/glossary#deterministic-computer-vision), with fixed rules and no trained models. ## How is accuracy measured, egg by egg? Accuracy is counted per egg: for each egg in a scene, the class given by Vigiovo is compared with the class in the ground truth. Accuracy is not measured per image, because an image with many correct eggs can still hide a cracked egg graded as clean. On video, Vigiovo follows each egg between frames and measures it three times; the decision uses a [two-of-three vote](https://vigiovo.com/en/resources/two-of-three-voting). In this mode, 376 of 381 tracked eggs were classified correctly, which is 98.7%. On a single frame, each image is judged on its own, and accuracy per egg was about 95%. **Accuracy measured on the synthetic scenes with ground truth** | Measure | Result | Conditions | | --- | --- | --- | | Per-egg accuracy, tracked video | 98.7% (376 of 381) | Synthetic scenes with ground truth | | Per-egg accuracy, single frame | about 95% | Synthetic scenes with ground truth | | Clean eggs rejected | 0 | All synthetic tests | ## What do these numbers not prove? They do not prove accuracy on your line. A synthetic scene is an approximation: it does not reproduce all the variation of a real conveyor, such as the particular shell colour of your batch, dust build-up, a lamp reflection on one part of the belt or camera vibration. The number shows that the measurement logic works on the scenes tested, and nothing more. A surface camera also cannot see inside the egg, with or without synthetic testing. The air cell, the yolk, internal blood and [hairline cracks](https://vigiovo.com/en/resources/glossary#hairline-crack) remain out of reach. For the glare that can turn into false dirt, see [glare and false dirty grades](https://vigiovo.com/en/resources/glare-and-false-dirty-grades). ## Why is calibration with your own eggs still needed? Because Vigiovo's limits are fixed configuration values, and calibration is what fits them to your camera, your light and your type of egg. A synthetic test confirms that the rule works; only the eggs on your line show where it should sit. The crack limit, for example, is tuned when each line is calibrated and is not a legal standard. So the next step is a pilot with eggs labelled by a trained team. The [step-by-step calibration guide](https://vigiovo.com/en/resources/calibrating-with-labelled-eggs) explains how to tune the limits, and [how to run a camera inspection pilot](https://vigiovo.com/en/resources/running-a-camera-inspection-pilot) describes the test design on the line. ## How do you read a synthetic result without fooling yourself? Read the accuracy per egg, not only the average. A high average can hide a class that often gets it wrong, so, where the test allows, split the errors by class before drawing any conclusion. Repeat the test on the same scenes before you trust it. Because Vigiovo is deterministic, the repetition should give the same result; if it does not, the problem is in the test setup, not in the classification. See [reproducible inspection](https://vigiovo.com/en/resources/reproducible-inspection). Keep a record for each egg. A result is only useful for audits if you can point, egg by egg, to the class given and the reason, as in [per-egg records for audits](https://vigiovo.com/en/resources/per-egg-records-for-audits). ## Frequently asked questions ### Does synthetic data prove that the system works in my plant? No. It shows that the measurement logic gets the tested scenes right, where the ground truth is known. To learn what happens at your plant, you need to measure with eggs and light from your own line, in a calibrated pilot. ### What does 98.7% accuracy mean? On the synthetic scenes, Vigiovo classified 376 of 381 tracked eggs correctly on video, which gives 98.7%. That figure comes from scenes generated with ground truth, not from a production line, and it does not cover internal defects the camera cannot see. ### Why are the video and single-frame accuracy figures different? The two figures measure different situations. On video, each egg is measured three times and the decision uses a two-of-three vote. On a single frame, the image is judged on its own, without that vote across measurements of the same egg. ## Sources - [Como o Vigiovo classifica ovos sem IA / How Vigiovo grades eggs without AI](https://vigiovo.com/como-funciona), Vigiovo ## Read next - [Integrating egg grading data with your systems](https://vigiovo.com/en/resources/integrating-egg-grading-data.md) (Guide, 2026-10-09): How to use an egg grader's per-egg records (class, scores, reasons, measurements) in dashboards and integrations, and what the data cannot tell you. - [Reproducibility: same image, same result](https://vigiovo.com/en/resources/reproducible-inspection.md) (Explainer, 2026-10-09): Why reproducibility matters for audits and disputes in egg grading, what makes it possible in Vigiovo and what it does not guarantee. - [Candling vs computer vision: what each one detects](https://vigiovo.com/en/resources/candling-vs-computer-vision.md) (Comparison, 2026-10-09): Candling compared with camera inspection in egg grading: which defects each detects, speed, record keeping and why the two methods work best together. ## How Vigiovo helps Vigiovo grades eggs on the conveyor as clean, dirty, cracked or unverified, in real time, with deterministic computer vision: every decision comes with the measurements behind it, on an ordinary CPU. https://vigiovo.com/en/how-it-works.md 4 min read.