Exploiting temporal context in 2d to 3d human pose regression

Varga, Viktor and Véges, Márton: Exploiting temporal context in 2d to 3d human pose regression. Conference of PhD Students in Computer Science, (11). pp. 169-172. (2018)

Cikk, tanulmány, mű cscs_2018_182-185.pdf Download (269kB)

Abstract

A major drawback of end-to-end image to 3d pose estimation approaches is the abscence of rich, in-the-wild image datasets with 3d human pose annotation. In this paper we show, that splitting the task and solving the subproblems of image based 2d pose estimation and 2d-to-3d coordinate regression independently is a viable approach. What is more, we present a lightweight deep learning based model to perform 2d-to-3d human body pose regression that is able to exploit temporal information and thus improve the state of the art.

Actions (login required)

View Item