Canonical Surface Mapping and Uncertainty Quantification in Deep Learning for 6D Pose Estimation
MetadataShow full item record
In this work, we focus on the 6D pose estimation problem for single-view images. We divide our thesis into two main parts. For the first part, we work on solving the pose estimation problem by extending the ACSM neural network to RGB-D images. Using the depth of the images, we create a novel algorithm called ACSM-D that can encode and learn from these depth information. Regarding the second part of our thesis, we work on building a framework to quantify uncertainty for the 6D pose estimation problem. We extend the deep evidential regression method to the pose estimation problem space. Finally, we apply our uncertainty quantification framework to ACSM by introducing a novel algorithm called U-ACSM to measure the uncertainty of its predictions, and we produce a qualitative and quantitative study of this new algorithm.