Relational Graph Representation Learning for Predicting Object Affordances

We address the problem of affordance classification for class-agnostic objects considering an open set of actions, by unsupervised learning of object interactions,inducing object affordance classes. A novel qualitative spatial representation incorporating depth information is used to construct Activity Graphs which encode object interactions. These Activity Graphs are clustered to obtain interaction classes, and subsequently extract classes of object affordances. Our experiments demonstrate that our method learns object affordances without being scene- or object-specific.