grantword8 / BLV

Balancing Logit Variation for Long-tailed Semantic Segmentation
21 stars 0 forks source link

Questions about the segmentation performance of classes with a large number of samples #7

Closed jetitime closed 4 months ago

jetitime commented 5 months ago

Hi @grantword8 , Thank you for your contribution to solving the problem of imbalanced number of samples in different categories in segmentation tasks. I applied BlvLoss to the unsupervised domain adaptation segmentation task from Potsdam to Vaihingen. Indeed, IoU is significantly improved in categories with a small number of samples(Car and Clutter). However, the segmentation performance of the categories with a large number of samples has dropped significantly(Impervious surface from 80 to 60). I observed that the unsupervised domain adaptation segmentation task in your paper did not have such a problem. I am wondering what caused this problem, and hope to get your advice. Thank you very much!