Fig.1: Motivation for IWE. Left→right: (i) Raw predictions before thresholding
for two classes — a true positive Person (red) (bbox cluster size = 167, max score
= 0.14) and a false positive Dog (blue) (bbox cluster size = 45, max score = 0.15);
(ii) In ZS, all scores remain below the detection threshold, causing a missed detection;
(iii) Standard entropy minimization over-confidently sharpens scores, resulting in a dog
false positive; and (iv) IWE focuses updates on dense bbox clusters and suppresses
isolated boxes, producing only the correct person detection.