Primary supervisorQiuhong Ke
Human action recognition plays an important role in various applications. Existing works assume that the training and testing share a common pre-defined list of action categories. Given a new unseen action during testing, the existing model will simply assign a wrong action category from the pre-defined list. This greatly limits the applicability of existing methods for practical model deployment.
This project aims to identify novel/unseen actions that the model does not train during model inference. In particular, the model should recognise the known action class and identify the novel actions.