confusion about out-of-candidate prediction

[zaiquanyang]

Thanks for your enlightening work! Here, I am confused about the out-of-candidate prediction of segmentation model. As show in Fig 3 of the paper, the pseudo label y_hat is not incomplete cosnidering the lack of background. I guess that, when given the complete pseudo label , the "chair" region should be more likely to be predicted as background instead of "chair" (out-of-candidate prediction). If this is true, I am confused about the out-of-candidate prediction defined in the work. Can you make it clear that how out-of-candidate prediction appears ?

[clownrat6]

Thanks for your question 🤗:

Why isn't the "chair" prediction of Fig.3 "background"?

1. Before facing the sample showing in Fig.3, the network may already be perturbed by other noisy pseudo labels which provides "sofa" -> "chair" supervision signal so that the network predict "soft" to "chair" when facing this sample.
2. After training on sample showing in Fig.3, the network may occur the situation descriped by you ("sofa" -> "background"). But how the network is specifically disturbed by noisy labels is difficult to understand so that the behaviour of single sample is just a reference for better understanding out-of-candidate prediction.

[xixiaos]

I would like to know the code for your specific job. May I ask when the code can be uploaded?