IDS peak 2.8.0 / uEye+ firmware 3.33

The threshold is a value above which the inference result is assumed to be true and below which it is not true.

Neural networks do not make unique predictions. The neural network determines a confidence score in its prediction, but no uniqueness with respect to the result. A neural network will not make the statement: "This is an apple", but "This is to 99 % an apple" or "This is to 60 % an apple". Which statement is really still true for an apple, i.e. when the statement becomes true or untrue, is the task of the threshold. It forces the unique assignment, which can then be output as a result. For example, if the threshold is 75, then all predictions that have a confidence score ≥ 75 are treated as true. The lower you set the threshold, the greater the uncertainty and error rate will become. For example, a pear classified as an apple based on its similarity with a confidence score of 60 will be called an apple at a threshold of 60. However, if the threshold is 75, the pear is not classified as an apple.

After setting the threshold, check the reliability of the inference results. If borderline patterns in particular are detected incorrectly, this may be due to a threshold that is too low.