Class-Level Object Detection and Pose Estimation from a Single RGB Image Only

[Subtitle]
Thesis by: Edward Cornelius Krubasik
Advisor: Prof. Nassir Navab
Supervision by: Fabian Manhardt, Federico Tombari
Due date: 15.9.2018
For further information please contact: Fabian Manhardt

Abstract

2D Object Detection has seen some great advancements over the last years. For instance, detectors like YOLO or SSD are capable of performing accurate localization and classification on a large amount of classes. Unfortunately, this does not hold true for current pose estimation techniques, as they have trouble to generalizing to a variety of object categories. Yet, most pose estimation datasets are comprised out of only a very small number of different objects to accommodate for this shortcoming. Nevertheless, this is a severe problem for many real world applications like robotic manipulation or consumer grade augmented reality, since otherwise the method would be stronlgy limited to this handful number of objects. Therefore, we would like to propose a novel pose estimation approach for handling multiple object classes from a single RGB image only. To this end, we would like to extend a very common 2D detector i.e. Mask R-CNN[1], to further incorporate 6D pose estimation. Eventually, the overall architecture might also involve fully regressing the 3D shapes of the detected objects.

Literature

[1] Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick. "Mask R-CNN", ICCV 2017