Finding the pose of objects in images is important in many robot applications. Lately, Deep Learning has showed impressive results, especially in estimating pose from a 2D image. The robustness for industrial applications has not been evaluated though. Current live systems typically use 2.5D images/point clouds (stereo, time-of-flight) to obtain the 3D information necessary for the pose estimation. Another research question is therefore how Deep Learning performs with this kind of image input?
The task in this Master’s Thesis is to evaluate and improve on 6D pose estimation in 2D and 2.5D images for industrial applications. The goal is to find network architectures that can be trained based on a given CAD model only. Thus, the work also involves simulating 2D and 2.5D scenes which are then used for the Deep Neural Network training.
For more information, contact R&D manager Ola Friman (firstname.lastname@example.org) and apply below at "ANSÖKAN" before the 19th of October. Please notice: Application must include CV, Cover letter and LADOK (study results including courses).