Why shrink function need to init mask？

[xuyaojian123]

Why does mask not get a new mask also through F.grid_sample, but chooses to initialize mask to 1. I found that mask quickly becomes 0 during the training process. If my bbox remains unchanged, how can I train mask effectively? Thank you.

[daniel03c1]

Thank you for your interest. We just choose to reinitialize the mask after the shrinking process. However, I would encourage you to explore various settings, as different configurations might yield interesting results. Thank you.

[xuyaojian123]

Thank you!