Unbiased Scene Graph Generation with External Knowledge and Causal Analysis

Scene Graph Generation is a task in which descriptive depictions of images are generated. Majorindustry applications of scene graph generation include image captioning, Visual Question Answering and gaming industries to detect graphics. Extracting graph representation, basically triplets can be a challenging task in Computer Vision. The most prevalent problem in scene graph generation at present is severe training bias towards more frequently occurring entities and context that interferes with the actual content. In this work, we combine two already existing state-of-the-art methods. We bridge knowledge graphs with scene graphs to basically remove long tail distribution and add context to the process of scene graph generation and then use counterfactual analysis to remove the contextual bias introduced due to addition of external knowledge.