Scalable Theory-Driven Regularization of Scene Graph Generation Models

Several techniques have recently aimed to improve the performance of deep learning models for Scene Graph Generation (SGG) by incorporating background knowledge. State-of-the-art can be divided into two families: one where knowledge is incorporated model in a subsymbolic fashion, and another which maintained symbolic form. Despite promising results, both families face several shortcomings: first requires ad-hoc, more complex neural architectures increasing training or inference cost; second suffers from limited scalability w.r.t. size Our work introduces regularization technique injecting SGG that overcomes limitations prior art. model-agnostic, does not incur any cost at time, scales previously unmanageable sizes. We demonstrate our accuracy state-of-the-art models, up 33%.