Learning with semantic ambiguity for unbiased scene graph generation

Scene graph generation (SGG) aims to identify and extract objects from images and elucidate their interrelations. This task faces two primary challenges. Firstly, the long-tail distribution of relation categories causes SGG models to favor high-frequency relations, such as "on" and "in". Secondly, some subject-object pairs may have multiple reasonable relations, which often possess a certain degree of semantic similarity. However, the use of one-hot ground-truth relation labels does not effectively represent the semantic similarities and distinctions among relations. In response to these challenges, we propose a model-agnostic method named Mixup and Balanced Relation Learning (MBRL). This method assigns soft labels to samples exhibiting semantic ambiguities and optimizes model training by adjusting the loss weights for fine-grained and low-frequency relation samples. Its model-agnostic design facilitates seamless integration with diverse SGG models, enhancing their performance across various relation categories. Our approach is evaluated on widely-used datasets, including Visual Genome and Generalized Question Answering, both with over 100,000 images, providing rich visual contexts for scene graph model evaluation. Experimental results show that our method outperforms state-of-the-art approaches on multiple scene graph generation tasks, demonstrating significant improvements in both relation prediction accuracy and the handling of imbalanced data distributions.