Learning Debiased Representations for Long-Tailed Visual Recognition

Lam, Prudence

Student Work

Learning Debiased Representations for Long-Tailed Visual Recognition

Public Deposited

Real-world data often exhibit a long-tailed distribution, posing a challenge for classification models that are inherently biased towards higher frequency classes. Despite extensive research on learning unbiased classifiers, the issue of representation bias remains under-explored. Our observations show a negative correlation between class frequency and intra-class variance in feature space. We explore the use of data augmentation, specifically mixup and implicit semantic data augmentation (ISDA) in learning more uniformly distributed features. Moreover, we use the class-conditional statistics obtained from ISDA to fit linear discriminant analysis (LDA) directly on the features, which we hypothesize to be more robust than Softmax. Extensive experiments demonstrate the competitiveness of our framework on four long-tail benchmarks.

This report represents the work of one or more WPI undergraduate students submitted to the faculty as evidence of completion of a degree requirement. WPI routinely publishes these reports on its website without editorial or peer review.

Creator