Class correlation correction for unbiased scene graph generation

Mengnan Zhao, Yuqiu Kong, Lihe Zhang, Baocai Yin

Abstract: Highlights•We propose two debiasing transformations to exploit the dependency knowledge hidden in classes.•We introduce two biasing transformations to estimate the bias learned by the neural network.•The debiasing and biasing transformations fully alleviate the biased prediction problem in SGG.