Confused about unknown localization

[jinweiisgreat]

Hi Joseph, thank you for your great work! Here, I have a question about how to localize the unknown instance. We know that unknown candidate proposals are from top-k region proposals which do not overlap the ground truth proposals, right? But according to RPN loss function: when $p_i^*$ is 1, the region proposal participates in training. But for unknown objects, the $p_i^*$ will be 0, the proposal of unknown objects cannot participate in training, right? So, from my understanding, we just using anchors as proposals to localize unknown objects, right? If I have misunderstood, please correct me. Thank you!