Towards More Generalisable Compositional Feature Learning in Human‐Object Interaction Detection | Synapse