Model-based inexact graph matching on top of DNNs for semantic scene understanding

Jérémy Chopin, Jean-Baptiste Fasquel, Harold Mouchère, Rozenn Dahyot, Isabelle Bloch

Abstract: Highlights•A priori structural information is modeled by a graph built from annotated images.•Original graph matching method between a DNN output and the a priori information.•Genericity with respect to the nature of the structural information (hyperparameter).•Relevance evaluated on two datasets, for various DNNs and training dataset sizes.•Both code and data are freely available.