Dealing with the ambiguity induced by object symmetry in pose estimation tasks

Thesis by: Diego Martin Arroyo
Advisor: Prof. Nassir Navab
Supervision by: Fabian Manhardt, Federico Tombari
Due date: 15.8.2018
For further information please contact: Fabian Manhardt

Abstract

The task of 3D pose estimation from 2D camera views is currently a very popular research topic. However, most approaches do not take into account the fact that a single view can result in several equally valid outcomes due to symmetries in an object. This uncertainty can lead to inaccurate results and ignores useful information in the image. In this thesis we intend to explicitly model this pose ambiguity as a multiple hypothesis prediction problem reformulating existing single-prediction approaches [1]. The resulting model should also collapse to a single outcome when there is only one valid pose. Additionally, we want to estimate the potential symmetry axes of an object based on the predicted poses. The final pipeline will also include a object detection system and should work in real time on standard hardware.

Literature

[1] Christian Rupprecht, Iro Laina, Robert DiPietro^?, Maximilian Baust, Federico Tombari, Nassir Navab, Gregory D. Hager. "Learning in an Uncertain World: Representing Ambiguity Through Multiple Hypotheses", ICCV 2017