Improving 6D Pose Estimation Accuracy of Articulated Objects by Considering Physical and Visual Plausibility

Kurzfassung

Accurately estimating the 6D pose of objects is a crucial capability for many robotic applications, but it remains challenging, particularly in cluttered and partially occluded environments. While there exist a wide variety of methods for estimating the 6D pose of objects from an image, they are often lack robustness. In contrast to this, humans can robustly estimate the poses by using addition clues---such as context, physics, or known relationships between objects---to determine if a pose is plausible. For example, we know that an object is unlikely to intersect the surface of a table or float above it, something that current methods often cannot detect. This highlights the need for more robust pose estimation approaches that can incorporate similar cues to improve pose estimation in real-world scenarios. In this thesis, a method to both identify and correct implausible pose estimates for known objects is presented. The method uses additional knowledge of the scene in order to estimate the plausibility of each pose estimate. It uses a versatile framework of constraints to quantify which poses are plausible. Examples of constraints include visual and physical plausibility. With these constraints, relationships between object poses can also be expressed. One use case for this are articulated objects, they can be represented as parts connected with joint constraints. Given one or multiple initial pose hypotheses, the proposed method iteratively estimates the plausibility of existing hypotheses and generates new pose hypotheses with the goal of finding a pose with a higher plausibility and accuracy for each object. Knowing the confidence of pose estimates is helpful for downstream robotic tasks like grasping objects. The developed method for estimating plausibility is used to calculate task-specific confidence values for each estimated pose. By recognizing pose estimates which are might be too imprecise, a robot can attempt to collect more information (e.g., look from a different perspective) before performing an action which requires an accurate pose. The method is evaluated on multiple datasets for both rigid and articulated objects with respect to the accuracy of both the pose estimates and their confidence.