In the “YOLO Algorithm” lecture by Andrew (Week 3), he notes that the target label’s 4th dimension is 5 + the # of classes, with the 5 being {p_c, b_x, b_y, b_h, b_w}. However, if the anchor box information and thus b_h and b_w are already encoded into the 3rd dimension, wouldn’t it be more practical to only put p_c, b_x, and b_y in the 4th dimension?
In other words, isn’t it redundant to include the height and width of the bounding box in the 4th dimension, if that information is already encoded in anchor boxes?
Anchor boxes and bounding boxes, either ground truth bounding boxes or predicted bounding boxes, are not the same things. Anchor boxes are abstract concepts, they represent shapes representative of object shapes in the training data set, but not the shape any one particular object from any particular image. Also, they have only shape, no location. Ground truth and predicted bounding boxes, on the other hand, have both location and shape because they represent specific objects in specific input images. The b_h and b_w values in the prediction output vector are bounding box shape not anchor box shape.
That clears it up – thank you!