Confidence Scores and Thresholds

Confidence

Part of: Computer Vision Basics

A vision model rarely says "cat." It says "cat, 0.93." That number is a confidence score , and the cutoff you pick for it quietly decides whether your system cries wolf or sleeps through a fire. Tune it one way and you flood operators with false alarms; tune it the other and you miss the thing you built the system to catch. Confidence and thresholds are where a vision model meets the real world. What it is A confidence score is the model's estimated probability for a prediction, a number between 0 and 1 . "Helmet detected, 0.97" means the model is very sure; "0.51" means it is barely past a coin flip. To turn that continuous score into a yes/no decision, you pick a threshold t: predictions at or above t are accepted as positive, everything below is treated as negative. The threshold is your dial, not the model's. The same model becomes trigger-happy at t = 0.3 and cautious at t = 0.9. Nothing about the model changes, only where you draw the line. How it works Every prediction, once you apply a threshold and compare it to the truth, lands in one of four buckets: - True positive (TP): predicted yes, actually yes. A real catch. - False positive (FP): predicted yes, actually no. A fals

Challenge: Tune the Threshold