I’m working on a project to segment rooftop objects, I have a labeled dataset for YOLOv8 from roboflow.
I’m creating a custom model using PyTorch, but I’m stuck as to how PyTorch gets the labels for instance segmentation.
Because it’s about objects on a roof, the labels I have are overlapping, here is an example:
"0 0.8151095629770003 0 0.7230230457121672 0.06468577367652012 0.629497676615071 0.1376042821845973 0.9193334366944352 0.4491040183592357 1 0.3833300151572437 1 0.16354047283849452 0.8438865996222608 0
0 0.5468747017258877 0.04966784730650573 0.41719574128953224 0.12416961826626435 0.35124472712475713 0.17141464375294052 0.7402872243045079 0.5745475758597562 0.8066917261524205 0.5246918030069853 0.9188415514955619 0.44950608104353224
0 0.2841732368719464 0.208170580740801 0.16228395620321057 0.29134432383450326 0.23119906089224523 0.2986127892939919 0.571328448551029 0.6468934258944896 0.5431665666792898 0.6944899882906383 0.5597133627503145 0.706839090537974 0.6625912687571204 0.6121626399750674 0.7316561567057452 0.5674706508894717 0.3505037044936923 0.17323176011781283"
the format is in class, x y coordinates, what is the format PyTorch need to do instance segmentation on this?
Online I found that it’s usually transformed into a mask with 0 as background, class 1 as 1, class 2 as 2, etc. how does that work when an overlapping class has a large overlap with a different class?
In this case, the orange class overlaps for the most part with the pink class (the orange class is on top of the pink class)