Question 819

Two Boxes, One Threshold

A detector proposes one box for an object. Coordinates are (x1,y1,x2,y2)(x_1, y_1, x_2, y_2) in pixels, top-left origin.

box coordinates area
ground truth (0, 0, 100, 100) 10000
prediction (20, 20, 120, 120) 10000

Under the PASCAL VOC detection rule, how is this prediction scored?

A

IoU is about 0.47, so the box is simply left unmatched and discarded, which lowers recall but leaves precision untouched.

B

Intersection 6400, union 13600, so IoU is about 0.47: below the 0.5 threshold, and the detection counts as a false positive.

C

The prediction covers 6400 of the ground truth's 10000 pixels, an overlap of 0.64, which clears the 0.5 threshold and counts as a true positive.

D

Intersection 6400 over the two areas summed, 20000, gives 0.32: below threshold, and the detection counts as a false positive.