[cs-talks] IVC Seminar, Jacopo Cavazza, Friday March 31st, 1pm @ Hariri Seminar Room (MCS 180)
Harrington, Jacob Walter
jwharrin at bu.edu
Tue Mar 28 14:49:12 EDT 2017
Intertwining Kernel Methods and Feature Learning for Covariance-Based 3D Action Recognition
Jacopo Cavazza, PhD Student, Istituto Italiano di Tecnologia – Genova
Friday March 31st, 1pm – 2pm
Action Recognition (AR) from skeletal joints is a popular research field due to the diffusion of depth sensors and motion capture systems which circumvent most issues (light variations, occlusions, ...) of video-based approaches. In 3D AR, kernel-based methods and neural networks are two antithetic state-of-the-art paradigms, seldom interconnected. In my talk, I’ll bridge this gap by proposing two novel methods.
1. KERNELIZED COVARIANCE – a new generalized covariance (COV) representation to implement a fully kernelized network of symmetric and positive definite matrices. This boosts action recognition since modelling arbitrary, non- linear relationship conveyed by the skeletal joints.
2. KRONECKER-GAUSS FEATURE MAPS – while approximating the log- Euclidean kernel between COV operators, we achieve an explicit and data-driven representation using a shallow neural network. This guarantees top performance, scalability to big data and computational efficiency for training/inference on CPU.
Jacopo Cavazza was born in Novi Ligure, Italy, in 1990. He graduated cum laude from the University of Genova, Faculty of Applied Mathematics, in 2014. Afterwards, he joined PAVIS department, where he is currently acting as PhD student, advised by Prof. Vittorio Murino. His research interests span image classification, crowd analysis, humans’ action recognition and intention prediction.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cs-talks