Unbiased Scene Graph Generation Based on Relationship Label Division

Scene Graph Generation (SGG) captures triplets <subject, predicate, object> in images, which is crucial for in- depth image understanding and AIoT (Artificial Intelligence of Things) applications. Existing scene graph datasets contain relationship labels with varied semantic information, causing directly trained models to inaccurately infer object relationships. For example, models tend to predict more frequent relationships like “on” rather than semantically detailed relationships like “walking on.” To address this, we classify relationship labels into spatial relationships (e.g., “on”) and visual relationships (e.g., “walking on”). We design two models, Spatial Relationships SGG (S-SGG) and Visual Relationships SGG (V-SGG), to predict these categories separately. By providing the models with different location and visual information inputs, they can focus on predicting the relationship labels they are responsible for. We then merge the results predicted by the two models to generate the final scene graph. Our approach effectively reduces long-tail bias and enhances several metrics on the standard Visual Genome benchmark.