Confidence Score

When you use something like YOLO, it gives out an estimate score about how confident it is about the score. That is the confidence score.

I thought it was just the softmax?

I thought that it was just when you apply Softmax on the output classes, which class it should be is the end value.

• Yes, that is correct, but for object detection, that is not the full picture.

Confidence = P(class and object) = P(object) × P(class | object) 

• P(object) → that is the Objectness Score
• P(class | object) = softmax(class logits)

Example At inference, the final class score for each class c is:

P(object and class=c) = P(object) × P(class=c | object)

So for a cat in a box:

• If P(object) = 0.9
• And P(class='cat' | object) = 0.7
• Then YOLO says P(cat and object) = 0.63 for that box.