Hey everyone, I’m kind of a latecomer to this challenge, and I honestly can’t figure out how you guys are using the geometry column.

The given masks don’t overlay perfectly on the images, and because of that, I couldn’t properly train a supervised model. Even when I tried, the results were pretty poor.

So I decided to go unsupervised completely ignored the geometry column and experimented with different techniques. I managed to reach an IoU of around 64, but I just can’t seem to get past that.

Can anyone explain how you’re using the geometry data column for training? Are you aligning the masks somehow or doing special preprocessing that I might’ve missed?