How are the assumed thresholds for Amazon Rekognition Custom Labels computed?

Question

In the **Per label performance** section of my model's evaluation summary in the Amazon Rekognition console, I can see an **Assumed threshold** metric. How are those metrics computed?

When I collected all the confidence levels computed for a given class in the test dataset, I could see that the minimum confidence level for my **KO** label was 0.87, while the minimum confidence level for the **OK** label was around 0.30. Is there some kind of heuristics going on (for example, 90% of the minimum confidence level observed for a class?), or is there a precise mathematical definition behind the assumed thresholds?

Also, is there a way to get the assumed thresholds programmatically with the Amazon Rekognition Custom Labels API?

Accepted Answer

The assumed threshold for each label is the value above which a prediction is counted as a true or false positive. The metric is calculated based on the best F1 score achieved on the test dataset during model training. For more information, see the Assumed threshold section in the Rekognition Custom Labels guide.

To get the assumed thresholds programmatically, use the Rekognition SDK. The assumed thresholds are listed in the summary file output.

To get the assumed thresholds programmatically, use the [Rekognition SDK](https://docs.aws.amazon.com/rekognition/latest/customlabels-dg/tr-metrics-api.html). The assumed thresholds are listed in the [summary file output](https://docs.aws.amazon.com/rekognition/latest/customlabels-dg/tr-summary-file-api.html).

How are the assumed thresholds for Amazon Rekognition Custom Labels computed?

Contenido relevante