We just mentioned that an anchor box "well" surrounds the dog in the image. If the ground-truth bounding box of the object is known, how can "well" here be quantified? Intuitively, we can mea- sure...

Urgent

Extracted text: We just mentioned that an anchor box "well" surrounds the dog in the image. If the ground-truth bounding box of the object is known, how can "well" here be quantified? Intuitively, we can mea- sure the similarity between the anchor box and the ground-truth bounding box. We know that the Jaccard index can measure the similarity between two sets. Given sets A and B, their Jaccard index is the size of their intersection divided by the size of their union: |AnB |AUB J(A, B) = (13.4.2) In fact, we can consider the pixel area of any bounding box as a set of pixels. In this way, we can measure the similarity of the two bounding boxes by the Jaccard index of their pixel sets. For two bounding boxes, we usually refer their Jaccard index as intersection over union (loU), which is the ratio of their intersection area to their union area, as shown in Fig. 13.4.1. The range of an IoU is between 0 and 1: 0 means that two bounding boxes do not overlap at all, while 1 indicates that the two bounding boxes are equal. Intersection loU = Union Fig. 13.4.1: IoU is the ratio of the intersection area to the union area of two bounding boxes.