Towards Efficient Human Action Retrieval Based on Triplet-Loss Metric Learning

Recent pose-estimation methods enable digitization of human motion by extracting 3D skeleton sequences from ordinary video recordings. Such spatio-temporal skeleton representation offers attractive possibilities for a wide range of applications but, at the same time, requires effective and efficient content-based access to make the extracted data reusable. In this paper, we focus on content-based retrieval of pre-segmented skeleton sequences of human actions to identify the most similar ones to a query action. We mainly deal with the extraction of content-preserving action features, which are learned using the triplet-loss approach in an unsupervised way. Such features are (1) effective as they achieve a similar retrieval quality as the features learned in a supervised way, and (2) of a fixed size which enables the application of indexing structures for efficient retrieval.