How did you evaluate the model with the best corr-coeff?

[maaft]

I was wondering if you first trained the Mask-Prediction-Head until you reached the best AP and afterwards the IoU-Prediction-Head with all other model parameters fixed?

Or did you incorporate the new scores into the AP calculation and therefore implicitly saved a "good" IoU-Prediction-Head?