Understand how an existing model behaves

Once a model has been trained and deployed, its predictions start appearing alongside your images. This guide explains how to use EyVz to understand what the model gets right, what it gets wrong, and where it is uncertain.

Note

A model is a tool that offers suggestions. Your role is to evaluate those suggestions honestly. EyVz is designed to expose uncertainty, not to hide it.

What you are looking for

When reviewing a model, you want to answer three questions:

1. Where does the model agree with human judgment? These are cases where the prediction matches what an annotator would choose.
2. Where does the model disagree? These are errors — the model predicts one thing, but the correct answer is something else.
3. Where is the model uncertain? These are predictions with low confidence, where the model is effectively guessing.

Reviewing predictions on the Home page

1. Open the Home view and set appropriate time and camera filters.

2. Show only predicted images. Set Show / hide predicted to “Predicted only” and click Apply filters.

3. Scan the image grid. Each card shows predictions below the image thumbnail as label : value [confidence%]. For object detection models, bounding boxes are drawn directly on the image.

4. Look at confidence scores. High confidence (for example, 90% or above) means the model is fairly certain. Low confidence (below 70%) means it is not sure. Use the confidence min/max filter to isolate images in different confidence ranges.

5. Compare predictions to annotations. For images that have been annotated, both the human label and the model prediction are shown on the card. Disagreements between the two indicate model errors.

Using filters to find problems

Low-confidence images

1. Set Confidence max to a value like 70.
2. Click Apply filters.
3. Browse the results. These are images where the model is least sure. Review them carefully — they often reveal what the model struggles with.

Tip

Low confidence does not always mean the model is wrong. It means the model is uncertain. Some of these images may be genuinely ambiguous — even to a human.

Mismatches between predictions and annotations

1. Set Show / hide annotated to “Annotated only” and Show / hide predicted to “Predicted only”.
2. Click Apply filters.
3. Examine cards where the prediction and annotation differ. Open these images in the annotation view for a closer look.

Filtering by specific prediction values

Use the Predictions filter to find all images where the model predicted a specific label and value. For example, enter Defect:Scratch to see every image the model classified as having a scratch. Then check whether those predictions look correct.

Use negation (!Defect:Scratch) to find images where the model did not predict a scratch — helpful for checking whether the model is missing real defects.

Adjusting the display threshold

The Only show prediction above threshold slider in the filter pane controls which predictions are shown on image cards. This is a display filter — it does not change the data, only what you see.

• Set it low (for example, 10%) to see all predictions, including uncertain ones.
• Set it high (for example, 90%) to see only the predictions the model is most confident about.

Note

This display threshold is separate from the annotation confidence threshold configured in the Thresholds Statistics view. The annotation threshold determines which predictions are offered as suggestions during annotation. The display threshold only affects what you see on image cards in the Home view.

Opening images for closer inspection

Click any image thumbnail to open it in the annotation view (read-only). In this view you can:

• See the full-size image with zoom and image filters
• Toggle between Annotation and Prediction display modes
• Examine bounding boxes and confidence scores in detail

Building a mental model

After reviewing a range of images across different confidence levels, you should be able to say:

• “The model is reliable for X” — high confidence, consistently correct
• “The model struggles with Y” — frequent errors or low confidence on certain types of images
• “The model is uncertain about Z” — low confidence in a specific area, worth watching

This understanding helps you decide whether to trust the model, where to focus your annotation effort, and whether the confidence thresholds need adjusting.

What’s next?

Thresholds Statistics Adjust confidence thresholds based on what you learned.

Compare model versions Evaluate whether a new model version is better.

Annotate from scratch Start labeling when no model exists yet.