Presentation in LAAS-CNRS / ANITI, 2020 Sep 24.
In this talk, Yann Labbé presents his recent ECCV’20 work on 6D object pose estimation “CosyPose: Consistent multi-view multi-object 6D pose estimation”.
He start by presenting our deep learning-based approach for 6D object pose estimation from a single RGB image. This method achieves state-of-the-art results on multiple benchmarks and was the winner of the BOP challenge 2020 at ECCV 2020. Despite great numerical performances this method is still inherently limited by its single-view nature and naturally fails in cluttered scenes, for example due to occlusions. To address these failure cases that are critical to the deployment of real robotic systems, he is proposing to use multiple images captured from noncalibrated cameras.
He then presents an extension to multi-view approach which is able to automatically process noisy or incomplete visual information from multiple cameras into a complete object-level scene interpretation and significantly improve the accuracy of the pose estimates and the robustness with respect to the failure modes inherent to single-view.
The talk is briefly concluded by a presentation of a current work in progress which aims at extending this method for rigid objects to articulated objects with known kinematic constraints such as robots.
15 окт 2020
