Statistics

The Statistics view gives you a visual overview of how predictions and annotations are distributed across your data over time. It helps you understand patterns, spot shifts, and monitor the overall state of your inspection data.

UI walkthrough

The page is divided into two areas:

Area	Description
Filter pane (left)	The same filter controls as the Home view. Use them to narrow the time range, rigs, cameras, or other criteria for the charts.
Charts (right)	A set of charts displaying different aspects of your data.

Available charts

Depending on your site configuration, the following charts may be available:

Chart	What it shows
Classification labels over time	How classification prediction labels are distributed across time buckets. Select a model from the dropdown to focus on a specific model’s predictions.
Object detection labels over time	How detected object types are distributed over time. Select a model to focus on its detections.
Number of images	Total images, annotated images, and triggered images over time.
Image annotation distribution	A stacked bar chart showing how annotations are distributed across labels over time.

Tip

Use the model selector dropdown above each chart to focus on predictions from a specific model. This is especially useful when multiple models are deployed.

Chart controls

Each chart may include controls to adjust its display:

Control	What it does
Y max / Y min	Set upper or lower bounds for the Y axis.
Chart type	Toggle between different chart types (line, bar).
Percentage view	Show values as percentages instead of absolute counts.
Accumulated view	Show cumulative totals over time instead of per-period counts.

Note

Not all controls are available for every chart. The controls visible depend on your site’s statistics configuration.

Exporting data

The Export Prediction Accuracy (CSV) button at the bottom of the page exports prediction accuracy data as a CSV file. This allows you to analyse the data in external tools such as spreadsheets.

How to use statistics effectively

Statistics are most useful when you look at them with a specific question in mind:

• Is the model improving over time? Check whether the proportion of correct predictions increases across recent time periods.
• Are certain labels consistently wrong? Look for labels that appear frequently in predictions but rarely in annotations — or vice versa.
• Has something changed in production? A sudden shift in label distribution may indicate a change in the product or the inspection setup, not a model problem.

Caution

Statistics reflect whatever data matches your current filters. A narrow filter range may produce misleading charts. Make sure the time range and other filters are broad enough to be representative.

What’s next?

Runs Compare model versions with structured comparison runs.

Thresholds Adjust the confidence thresholds that affect how predictions are evaluated.

How-To: Evaluate a model A step-by-step guide to investigating model behaviour.