Twice the Picture, Same Network

A 16-class classifier is trained on $224\times224$ images. Its backbone is five stages, each of which is some "same"-padded $3\times3$ convolutions followed by $2\times2$ max pooling with stride $2$; the final stage has $512$ filters.

Two different heads were tried on that identical backbone:

• Head A — flatten the final block, then one dense layer down to the $16$ class scores.
• Head B — global average pooling, then one dense layer down to the $16$ class scores.

Both were trained to about the same accuracy and saved to disk.

A year later the camera feeding the system is replaced and the pictures arriving are $448\times448$. Nobody retrains anything and nobody resizes anything — the bigger pictures are fed straight into the two saved networks.

What happens when each one is run on a $448\times448$ image?