One class, three puppies, five detections. Drag τ and everything moves together: a detection scoring below the threshold fades from the photograph, the counts are recomputed over what is left, and the operating point walks along the curve on the right. The curve itself never moves — it is the whole sweep, and the threshold only chooses which of its points you are standing on. Watch what the two ends cost. At τ = 0 every detection is kept, all three puppies are found and recall is 1, but two of the five are wrong. Push τ past 0.10 and puppy 3 is lost: recall drops to 2/3 and never recovers, because raising a threshold can only ever remove detections. There is no true-negative count. A classifier has one because it makes a decision about every image; a detector does not, because there is no enumerated list of boxes it declined to predict, so precision and recall are the only two rates available and accuracy is not defined here at all.