This is the 'half-time'-seminar for PhD student Ivar Persson.

Abstract

6DoF pose estimation is the method of estimating an objects position, in real world coordinates, along with its rotation, while 3D shape reconstruction concerns the reconstruction of the actual shape of the object. Robust and accurate 3D shape reconstruction and 6DoF pose estimation are crucial for the development of (semi-)autonomous robots, including self-driving cars. This thesis aims to develop these methods in the monocular RGB camera setting, i.e. one ordinary camera. The usage of multiple cameras enable triangulation of objects and RGB-D cameras provide a notion of depth with every pixel. While useful, these methods are either more cumbersome (due to calibration requirements between the cameras in the multi camera use-case) or more expensive (in the case of RGB-D cameras).

This half-time seminar will dive into the methods of 3D reconstruction and pose estimation, both the classical methods and the newer that require Deep Neural Networks. We will discuss common datasets as well as how they have been used and what requirements there are on a dataset in order to be used with the methods planned to be used in the thesis. The current state-of-the-art will be showed before we discuss the three papers I have been involved in thus far. Paper I, Seg2Pose takes a segmentation mask as input and estimates the pose of a road user with one or multiple cameras, while Paper II, GUTS, also tracks the road user in 3d world coordinates. Paper III introduces a tighter shape of the road user with the help of learnt shape parameters and an autoencoder. Finally, a big portion will be earmarked for the current and future works and planned activities.