Compositional Zero Shot Learning

Compositional Zero-Shot Learning (CZSL) is a task in the field of computer vision that aims to recognize unseen state-object compositions during training. The core challenge of this task lies in the intrinsic entanglement of states and objects in images, requiring the model to generalize to novel combinations. Evaluation metrics for CZSL include accuracy on seen and unseen compositions as well as their harmonic mean (HM), making it valuable for applications such as cross-domain object recognition and scene understanding.