Every position of the feature map carries the same k = 9 boxes — three scales by three aspect ratios, all centred there — so the anchor count is the grid size times nine. Work it out for stride 16 before reading it off, then see how many of those anchors hang off the edge of the image and are dropped before training. Drag over the grid to draw one position's nine anchors on the image, to scale. With the scales of the Faster R-CNN paper nothing reaches the 0.70 threshold on either object — on the threshold alone both boxes would match nothing and train nothing, which is why the highest-IoU anchors are made positive whatever their IoU. The dog's box has the area of a 334 px square, so try scales stepping by ∛2 from 256 — 256, 323, 406 — and ordinary positives appear.