Research Project

EGOPOSE: 3D Human Pose Estimation from Ego-Camera Images


UPC Project

Start Date


End Date


Project Code


Default project illustration


Project Description

This is a project within the Call for IRI MdM Internal Projects 2018, a call for R&D projects addressed to IRI early career researchers, under the María de Maeztu Unit of Excellence Programme.

In this project we propose to estimate the 3D pose of a human from ego-camera images. In concrete, we want to focus on a scenario when a person or robot (ego-camera) is interacting with another person, therefore they are very close to each other. This kind of situation is very challenging because the images recorded will contain many occlusions and some body parts will not be visible. We plan to base our algorithm design on a generative adversarial network (GAN) paradigm composed of a generative network that will estimate 2D, 3D and camera poses, and a discriminative network that will try to guess if estimated parameters are plausible or not. To train this network we will create a new synthetic dataset with images rendered from a realistic human models and scenarios. Finally, we plan to compare quantitatively and qualitatively our work with state-of-the-art algorithms on 3D human pose estimation.