2019/06/13 by Suhas Lohit, Qiao Wang, Lohit, Suhas +3 · 1 citation
Computer Science · #Anomaly Detection Techniques and Applications #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Human Pose and Action Recognition #Time Series Analysis and Forecasting
paper · pdf · doi:10.48550/arxiv.1906.05947
openalex publication_date 2019/06/13 · openalex created_date 2022/07/28 · openalex updated_date 2026/07/28
Many time-series classification problems involve developing metrics that are\ninvariant to temporal misalignment. In human activity analysis, temporal\nmisalignment arises due to various reasons including differing initial phase,\nsensor sampling rates, and elastic time-warps due to subject-specific\nbiomechanics. Past work in this area has only looked at reducing intra-class\nvariability by elastic temporal alignment. In this paper, we propose a hybrid\nmodel-based and data-driven approach to learn warping functions that not just\nreduce intra-class variability, but also increase inter-class separation. We\ncall this a temporal transformer network (TTN). TTN is an interpretable\ndifferentiable module, which can be easily integrated at the front end of a\nclassification network. The module is capable of reducing intra-class variance\nby generating input-dependent warping functions which lead to rate-robust\nrepresentations. At the same time, it increases inter-class variance by\nlearning warping functions that are more discriminative. We show improvements\nover strong baselines in 3D action recognition on challenging datasets using\nthe proposed framework. The improvements are especially pronounced when\ntraining sets are smaller.\n