@roytseng-tw, hi, thanks for your excellent work, I just curious about the mask loss that in your implementation.
In the paper, it says that they use the predicted label by the classification to select the output mask, so when computing the loss, whether it should use the target label of this rois to select the predicted mask first, and then compute the binary loss, while in your code where hasn't the step of selecting and do binary loss directly. I wonder whether my understanding is right or not, could you give some appreciate advice?
@roytseng-tw, hi, thanks for your excellent work, I just curious about the mask loss that in your implementation. In the paper, it says that they use the predicted label by the classification to select the output mask, so when computing the loss, whether it should use the target label of this rois to select the predicted mask first, and then compute the binary loss, while in your code where hasn't the step of selecting and do binary loss directly. I wonder whether my understanding is right or not, could you give some appreciate advice?