Overfitting on "iteration" parameters?

[firrice]

Hi, the unify framework for all knids of paired-loss proposed in the paper is great, while i found that it appeared that the best "test recall" has been actually decided by val_dataset, which refenced to the raw code below:

According to the fig above, "val datatset" actually also plays a role of "test dataset", which means "test dataset" is visible during training.
So does it seems like choosing a "best train iteration" parameter, which is a risk of overfitting on training hyperparameters?
(I have found similar operation in several other papers, and i knew there was a lack of test dataset building the dataset, such as the general protocal "construct query+gallery based on the raw val+test split in DeepFashion")