Reviews: Pixels to Graphs by Associative Embedding

This paper proposes the use of a Hourglass net with associative embeddings to generate a graph (relating objects and their relationships) from an image. The model is presented as one of end-to-end learning. The Hourglass net provides heatmaps for objects and relationships, feature vectors are extracted from the top locations in the heatmaps and used with FC layers for predicting object-classes, bounding boxes and relationships among objects. Associate embeddings are used to link vertexes and edges, each vertex having an unique embedding. Experimental results show the high performance of the proposed methodology.