I was skimming through the code and saw that you convert the rotation target to an allocentric representation.
I can't seem to find this conversion in the PoseCNN paper or original implementation.
I also haven't seen any discussion of this in rotation regression literature.
Have you compared performance with and without this conversion?
Did you use the allocentric representation in the original implementation (paper results)?
Is there a paper discussing this, that I have missed?
Hi Yu Xiang, Great work!
I was skimming through the code and saw that you convert the rotation target to an allocentric representation. I can't seem to find this conversion in the PoseCNN paper or original implementation. I also haven't seen any discussion of this in rotation regression literature.
Have you compared performance with and without this conversion? Did you use the allocentric representation in the original implementation (paper results)? Is there a paper discussing this, that I have missed?
Best regards, Rasmus