Auxiliary points confidence and offset prediction

[Indraa145]

Hello, thank you for the great work. I see that the training code is incomplete, so I'm trying to implement it and have a few questions regarding auxiliary points confidence and offset prediction:

1. For the auxiliary points, did you use the same classification and regression head you used to predict proposal point confidence and offset?
2. If yes, how did you do this? Did you increase the number of point to be predicted in the classification and regression head? Or did you feed the network twice? Once for point proposal prediction, and once for auxiliary points prediction?
3. If no, did you perhaps use a separate regression & classification branch for auxiliary points, which is used only for training? During testing these separate branch is discarded then? But if so, then what's the use of it? To help supervise the Encoder and IFI decoder?
4. Or perhaps did you use another way? If so, can you please share it?

Thank you for your time, I really appreciate it.