Towards a Fair Evaluation of Zero-Shot Action Recognition Using External Data

Zero-shot action recognition aims to classify actions not previously seen during training. This is achieved by learning a visual model for the seen source classes and establishing a semantic relationship to the unseen target classes e.g. through the action labels. In order to draw a clear line between zero-shot and conventional supervised classification, the source and target categories must be disjoint. Ensuring this premise is not trivial, especially when the source dataset is external. In this work, we propose an evaluation procedure that enables fair use of external data for zero-shot action recognition. We empirically show that external sources tend to have actions excessively similar to the target classes, strongly influencing the performance and violating the zero-shot premise. To address this, we propose a corrective method to automatically filter out too similar categories by exploiting the pairwise intra-dataset similarity of the labels. Our experiments on the HMDB-51 dataset demonstrate that the zero-shot models consistently benefit from the external sources even under our realistic evaluation, especially when the source categories of internal and external domains are combined.