Description:

With the rise of deep learning, there has been great advances in monocular 6d object pose estimation [1]. However, current methods have several limitations, which prohibits them from being widely applied in real applications. For example, continuing to train new 3d instances from the trained network models will decrease its performance on the old 3d object models. Recently works utilize knowledge distillation for object detection tasks [2][3], however little work has been done for 6d object pose estimation in this field. Therefore, the aim of the thesis is to incrementally learn to detection 6d object pose without forgetting.

Requirement:

Students should be familiar with pytorch or tensorflow and have hands-on experience with deep learning in past projects. Aiming for a paper publication would be a plus.


LatentFusion[2]

Knowledge Distillation [3]

Benefits:

you can learn state of art 6d object pose estimation networks and cooperate with experts both at the Chair and inside our industrial corporate partners.

[1] Wang, Gu, et al. “GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

[2] Zhou, Wang, et al. “Lifelong Object Detection.” arXiv preprint arXiv:2009.01129 (2020).

[3] Wu, Yue, et al. “Large scale incremental learning.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.