Skew Class-Balanced Re-Weighting for Unbiased Scene Graph Generation

An unbiased scene graph generation (SGG) algorithm referred to as Skew Class-Balanced Re-Weighting (SCR) is proposed for considering the predicate prediction caused by long-tailed distribution. The prior works focus mainly on alleviating deteriorating performances of minority predictions, showing drastic dropping recall scores, i.e., losing majority performances. It has not yet correctly analyzed trade-off between and in limited SGG datasets. In this paper, alleviate issue, loss function considered models. Leveraged skewness biased SCR estimates target weight coefficient then re-weights more predicates better trading-off ones. Extensive experiments conducted standard Visual Genome dataset Open Image V4 V6 show generality with traditional