leonhlj / teach-detr Goto Github PK

Hello, I have several questions about the ablations study section.

1. what does it mean by online teacher vs. offline teacher? Does online teacher means you also train the teacher as well? How do you train the teacher, by tracking the EMA? If so, this is self-distillation?
2. For the soft label vs. hard label section, have you confirmed it's because of the softmax of the MaskRCNN that causes the drop in performance. any experiments to verify it?

Can the increase in performance be seen as a result of the increase in the Recall of the dataset's annotation, although that new annotation is noisy? Have you tried to train purely with the Aux boxes without GT. For instance, start with a H-DETR, train it regularly, and then train another H-DETR student using the outputs from the 1st model.