Propose an architectural change to your existing single-shot network that incorporates iterative refinement of estimates using a recurrent structure. Explain the tradeoffs and benefits of this design choice.
Modify the single-shot network to use a recurrent convolutional architecture, where each step refines the estimate of labels by using a previous estimate as input. This change improves accuracy and robustness against noise by allowing for iterative refinement of estimates, while also increasing computational complexity due to the additional recurrence steps.