2017/05/23 by Anoop Cherian, Suvrit Sra, Cherian, Anoop +3
Computer Science · Engineering · #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Generative Adversarial Networks and Image Synthesis #Human Motion and Animation #Human Pose and Action Recognition
paper · pdf · doi:10.48550/arxiv.1705.08583
openalex publication_date 2017/05/23 · openalex created_date 2022/10/01 · openalex updated_date 2026/07/28
Representations that can compactly and effectively capture temporal evolution\nof semantic content are important to machine learning algorithms that operate\non multi-variate time-series data. We investigate such representations\nmotivated by the task of human action recognition. Here each data instance is\nencoded by a multivariate feature (such as via a deep CNN) where action\ndynamics are characterized by their variations in time. As these features are\noften non-linear, we propose a novel pooling method, kernelized rank pooling,\nthat represents a given sequence compactly as the pre-image of the parameters\nof a hyperplane in an RKHS, projections of data onto which captures their\ntemporal order. We develop this idea further and show that such a pooling\nscheme can be cast as an order-constrained kernelized PCA objective; we then\npropose to use the parameters of a kernelized low-rank feature subspace as the\nrepresentation of the sequences. We cast our formulation as an optimization\nproblem on generalized Grassmann manifolds and then solve it efficiently using\nRiemannian optimization techniques. We present experiments on several action\nrecognition datasets using diverse feature modalities and demonstrate\nstate-of-the-art results.\n