Transductive Universal Transport for Zero-Shot Action Recognition

This work addresses the problem of recognizing action categories in videos for which no training examples are available. The current state-of-the-art enables such a zero-shot recognition by learning universal mappings from videos to a shared semantic space, either trained on large-scale seen actions or on objects. While effective, universal action and object models are biased to their seen categories. Such biases are further amplified due to biases between seen and unseen categories in the semantic space. The amplified biases result in many unseen action categories simply never being selected during inference, hampering zero-shot progress. We seeks to address this limitation and introduce transductive universal transport for zero-shot action recognition. Our proposal is to re-position unseen action embeddings through transduction, \ie by using the distribution of the unlabelled test set. For universal action models, we first find an optimal mapping from unseen actions to the mapped test videos in the shared hyperspherical space. We then define target embeddings as weighted Fr\'echet means, with the weights given by the transport couplings. Finally, we re-position unseen action embeddings along the geodesic between the original and target, as a form of semantic regularization. For universal object models, we outline a weighted transport variant from unseen action embeddings to object embeddings directly. Empirically, we show that our approach directly boosts universal action and object models, resulting in state-of-the-art performance for zero-shot classification and spatio-temporal localization.