Accuracy is a classification evaluation metric that measures the fraction of predictions a model gets right, the number of correct predictions divided by the total. It is the most intuitive metric, but it can be badly misleading on imbalanced datasets, which are the norm in visual AI.
Accuracy is the share of predictions a model gets correct: correct predictions divided by total predictions, reported from 0 to 1 or as a percentage. It applies to classification, where each prediction is either right or wrong. Its appeal is that anyone understands it instantly. Its weakness is that it treats every class and every error as equally important and ignores how the classes are distributed, which is exactly where it breaks down on real visual data.
Key takeaways
Accuracy is correct predictions over total predictions, from 0 to 1.
It is the most intuitive classification metric and a fine summary when classes are balanced.
On imbalanced data it is misleading, so pair it with precision, recall, F1, and per-class metrics.
Where it works and where it fails
Balanced classes, equal error costs: accuracy is a reasonable single summary.
Imbalanced classes, a rare defect or rare disease: a model that always predicts the majority class scores high accuracy while catching none of the cases that matter.
How it works, and how FiftyOne fits
FiftyOne's evaluation reports accuracy alongside precision, recall, F1, and a confusion matrix, and lets you slice accuracy by class or by subset so you can see where a strong headline number hides a weak spot. See the FiftyOne Evaluating Object Detections workflow to learn more.
Accuracy vs related metrics
Summary of accuracy metrics and when they're best used
Summary of accuracy metrics and when they're best used
Metric
What it captures
Best when
Accuracy
Overall fraction correct
Classes are balanced, errors cost the same
Precision
Of predicted positives, how many are right
False positives are costly
Recall
Of actual positives, how many are found
False negatives are costly, or the class is rare
F1
Harmonic mean of precision and recall
You need to balance both on imbalanced data
Why it matters
Accuracy is the metric everyone reaches for first, which is exactly why it misleads so often. Information-gain insight: the accuracy paradox. On a dataset that is 99% one class, a model that blindly predicts that class every time is 99% accurate and completely useless. In visual AI, imbalance is the default, most pixels are background, most frames contain no defect, so accuracy alone routinely certifies a model that fails at its actual job. Report per-class metrics and slice by the rare, important cases, because that is where accuracy quietly lies.
Frequently asked questions
What is accuracy?
The fraction of predictions a model gets right, correct predictions divided by the total.
Is high accuracy always good?
No. On imbalanced data a model can score high accuracy while missing the rare class that matters most.
What is the difference between accuracy and precision or recall?
Accuracy is overall correctness, precision and recall focus on how well a model handles a specific class and its false positives or false negatives.