Illustration of our final CNN backbone architecture. The left column provides the layer structural properties, while the right column indicates the spatial output dimension for each layer starting from a 256×256 input size. The input image is on the top, and layers are stacked in order vertically. The width of a layer represents its number of filters. The green color indicates a layer that preserves the spatial dimension, while the red indicates a reduction. The output grid dimension and the list of the predicted parameters for each box are also indicated.