The paper "Neuro-Symbolic Scene Graph Conditioning for Synthetic Image Dataset Generation", developed from Giacomo Savazzi's master's thesis - currently a PhD student in Information Technology at Politecnico di Milano’s AIRLab - and supervised by Eugenio Lomurno, Cristian Sbrolli, Agnese Chiatti, and Prof. Matteo Matteucci, has been awarded the Leonardo Fibonacci Best Paper Award at the International Conference on Machine Learning, Optimization, and Data Science.
The research, conducted entirely at Politecnico di Milano's AIRLab, investigated the use of neuro-symbolic approaches for synthetic image dataset generation. The work proposes an innovative method that integrates structured symbolic knowledge – in the form of graphs representing objects and their spatial and semantic relationships – into the image generation process, enabling the explicit encoding of relational constraints that are difficult to capture with conventional approaches.
The results demonstrate that synthetic data generated through this approach, while exhibiting lower perceptual quality when evaluated individually, provide complementary structural information that significantly enriches real datasets when used for augmentation. This enables performance improvements in evaluation metrics, surpassing traditional generation methodologies and opening new perspectives for addressing data scarcity even in complex visual reasoning tasks.