Don't merge just review the kfloss code!!!!!
The kfloss is composed of 2 losses: xy_loss(center point diff) and kf_loss(distribution diff). I divided two losses separately for printing purposes, but overall losses are calculated together, so there's no need to modify if u want to take a look and train.
From the original source code, it requires inputting xy loss for not decoded prediction and target, while the decoded version for kf loss, I'm still not quite sure why is that the case, please help me justify if u can come up with some notions.
Don't merge just review the kfloss code!!!!! The kfloss is composed of 2 losses: xy_loss(center point diff) and kf_loss(distribution diff). I divided two losses separately for printing purposes, but overall losses are calculated together, so there's no need to modify if u want to take a look and train. From the original source code, it requires inputting xy loss for not decoded prediction and target, while the decoded version for kf loss, I'm still not quite sure why is that the case, please help me justify if u can come up with some notions.