Hello Abdessalem,
The iterations here do not mean gradient updates, and iterative refinement does not imply fine tuning either. They imply the number of refinement steps, i.e. how many times the network is re-used conditioned on its previous result (all in a *single* gradient update). Each step can share the same network with the same parameters, not just the architecture. Anyway, the parameter *updates* themselves are not relevant to the question. See Slide #108 in the Delineation slides, and it should become clear!
Best.
The iterations here do not mean gradient updates, and iterative refinement does not imply fine tuning either. They imply the number of refinement steps, i.e. how many times the network is re-used conditioned on its previous result (all in a *single* gradient update). Each step can share the same network with the same parameters, not just the architecture. Anyway, the parameter *updates* themselves are not relevant to the question. See Slide #108 in the Delineation slides, and it should become clear!
Best.
