Causal Property Based Anti-conflict Modeling with Hybrid Data Augmentation for Unbiased Scene Graph Generation

Scene Graph Generation (SGG) aims to detect visual triplets of pairwise objects based on object detection. There are three key factors being explored to determine a scene graph: visual information, local and global context, and prior knowledge. However, conventional methods balancing losses among these factors lead to conflict, causing ambiguity, inaccuracy, and inconsistency. In this work, to apply evidence theory to scene graph generation, a novel plug-and-play Causal Property based Anti-conflict Modeling (CPAM) module is proposed, which models key factors by Dempster-Shafer evidence theory, and integrates quantitative information effectively. Compared with the existing methods, the proposed CPAM makes the training process interpretable, and also manages to cover more fine-grained relationships after inconsistencies reduction. Furthermore, we propose a Hybrid Data Augmentation (HDA) method, which facilitates data transfer as well as conventional debiasing methods to enhance the dataset. By combining CPAM with HDA, significant improvement has been achieved over the previous state-of-the-art methods. And extensive ablation studies have also been conducted to demonstrate the effectiveness of our method.