logit adjustment
A classifier rebalancing method for handling long-tailed distributions that compensates for imbalanced class priors by adjusting the logit values output by the model.
Timeline (last 90 days)
Long-tailed recognition typically adopts a two-stage training strategy: the first stage learns a shared feature encoder on the original imbalanced data, and the second stage freezes the encoder and balances the classifier through resampling, reweighting, or classifier correction (such as logit adjustment and τ-normalization).
Common strategies for dealing with long-tail distributions include resampling, loss function weighting (such as focal loss), classifier rebalancing (such as logit adjustment), and merging tail categories into superclasses.
All Facts (2)
Long-tailed recognition typically adopts a two-stage training strategy: the first stage learns a shared feature encoder on the original imbalanced data, and the second stage freezes the encoder and balances the classifier through resampling, reweighting, or classifier correction (such as logit adjustment and τ-normalization).
50%UnverifiedCommon strategies for dealing with long-tail distributions include resampling, loss function weighting (such as focal loss), classifier rebalancing (such as logit adjustment), and merging tail categories into superclasses.
50%