Elisa Ricci - Learning from noisy and missing data: introducing matrix completion for human behaviour analysis from visual inputs
While vital for handling most multimedia and computer vision problems, collecting large scale fully annotated data sets is a resource-consuming, often unaffordable task. In this context, developing methodological approaches able to deal with noisy and incomplete data sets is highly desirable. Matrix completion is a generic framework aiming to recover a matrix from a limited number of (noisy) entries.
In this talk I will present three recent approaches derived from the matrix completion paradigm to address different visual tasks in the area of human behaviour understanding. First, coupled matrix completion for the analysis of complex social scenes. Second, non-linear matrix completion to recognize emotions elicited from paintings. Third, self-adaptive matrix completion for remote heart-rate estimation from videos.