SSD stands for “Single Shot MultiBox Detector”.
Core idea
Generate a bunch of boxes to look at at different scale. For each of them, output class confidence + offset to the shape of the default box. Compute the loss by matching each box with ground truth box and do regression.
The classifier is a 3*3 conv kernel.
