The widget opens on an exercise. Blue is yours: drag it to move it, pull a corner to resize it, and put it where you would draw the box round the dog. Nothing is scored yet, because the answer is hidden. When you are satisfied, press show ground truth: the gold box drawn by hand for this book appears, your placement is scored against it, and that score is kept so you can see what your eye was worth before you started tuning. The shaded region is the intersection, the dashed envelope is the union, and IoU is the first divided by the second. Then ask the second question: how far can the blue box slide before it stops counting as a match at 0.50, and does that distance feel like the difference between a right answer and a wrong one? Keyboard: ← ↑ → ↓ move the box, shift with an arrow drags its bottom-right corner and alt with an arrow drags its top-left one. The detector's box is what Faster R-CNN actually returned for this dog, scoring 0.996. It sits within four pixels of the gold box on every edge. The gold box was drawn by hand for this book rather than taken from a labelled dataset, so read that agreement as a demonstration of what a high IoU looks like, not as a measurement of the detector.