Thank you for providing this fantastic work. I just had a few questions regarding your methodology:
How are you calculating mIoU? In the paper you say you merge children to construct the parent nodes but how is this done practically?
Do you represent your ground truth as HxWxC where C is mainly sparse? Or do you perform a mapping such that typical segmentation ground truth can be used? If so, how do you practically perform BCE?
I looked through the open source implementation but many codes are missing documentation/comments which has lead to challenges in interpreting the methodology.
Hello!
Thank you for providing this fantastic work. I just had a few questions regarding your methodology:
I looked through the open source implementation but many codes are missing documentation/comments which has lead to challenges in interpreting the methodology.
Thank you!