According to the paper SSD: Single Shot MultiBox Detector, for each cell in a feature map k boxes are acquired and for each box we get $c$ class scores and $4$ offsets relative to the original default box_shape. This means that we get $m \times n \times (c +4) \times k$ outputs for each $m \times n$ feature map.
However, it is mentioned that in order to train the SSD network only the images and their ground truth boxes are needed.
How exactly can one define the output targets then? What is the format of the output in the SSD framework? I think it cannot be a vector with the positions, sizes and class of each boundary box, since the outputs are a lot more and relate to every default box in the feature maps.
Can anyone explain in more detail how can I, given an image and its boundary boxes' info, construct a vector that will be fed into a network so that I can train it?