Abstract: The performance of current scene graph generation models is affected by training with cross-entropy loss, exacerbating the problem of prediction bias stemming from biased training data.