Confusion about final loss

[rjtshrm]

@zmurez-ml First of all, thanks for this amazing work. I am training the network with some modification however during training, the networks learns to output only 1's as the loss is dominated by free space. To solve this you mentioned in the paper that the loss is only back propogated through observed space (i.e., between 1 and -1) which you also used. However, I am confused that if this is the case and we mask the final output for target tsdf values b/w 1 and -1, how does network learn the free space as no loss is calculated here i.e., no flow of gradients.