Can you please explain how (in implementation) is ground truth matched to the default boxes? For example, after the Conv4_3 layer, how are the default boxes in each cell of the 38x38 feature map mapped to the ground truth given in terms of the image space which is much bigger?