Global Semantic Descriptors for Zero-Shot Action Recognition

The success of Zero-shot Action Recognition (ZSAR) methods is intrinsically related to the nature semantic side information used transfer knowledge, although this aspect has not been primarily investigated in literature. This work introduces a new ZSAR method based on relationships actions-objects and actions-descriptive sentences. We demonstrate that representing all object classes using descriptive sentences generates an accurate object-action affinity estimation when paraphrase as embedder. also show how estimate probabilities over set action only without hard human labeling. In our method, from these two global classifiers (i.e., which use features computed entire video) are combined, producing efficient knowledge model for classification. Our results state-of-the-art Kinetics-400 dataset competitive UCF-101 under evaluation. code available at https://github.com/valterlej/objsentzsar