Is it appropriate to select predicted mask based on the GT label ? Information Leak?

[Deleted user]

Dear author

In the subsequent step, you are choosing a mask by relying on the ground truth class index derived from the batch_data. This approach may inadvertently lead to information leakage during inference, as it involves evaluating label information. Could the author kindly provide further insights and commentary on this potential issue?
https://github.com/JiabenChen/iQuery/blob/315fba343ec240b63f8632f0347940efda5b4c51/code/models/maskformer_predictor.py#L505

Cheers,
Ryu