Fast 6D Object Pose Refinement via Implicit Surface Representation Driven Optimization

Bo Pang, Ge Gao, Zhihao Cui, Xiaolin Hu, Xiangyang Ji, Xianming Liu

Keywords: Signed Distance Field, 6D pose refinement, Implicit neural network, ICP

TL;DR: In this paper, we propose a simple yet efficient self-supervised point cloud aligenment method via implicit neural network, which can serve as an alternative of ICP to achieve fast and accurate pose refinement.

Abstract:

Pose refinement after the initial pose estimator has been demonstrated to be effective for 6D object pose estimation. The iterative closest point (ICP) is the most popular refinement strategy, which however suffers from slow convergence due to the nature of iterative nonlinear optimization. In this paper, we propose a simple yet efficient self-supervised point cloud aligenment method via implicit neural network, which can serve as an alternative of ICP to achieve fast and accurate pose refinement. Our key idea is to encode the surface of target point cloud into a signed distance function (SDF); the optimal rigid transformation then can be derived by addressing a minimization problem over the SDF. The workflow of our method does not require any pose annotations. Experimental results show our method can achieve 6.4\%, 16.2\%, and 3.9\% performance improvement over the prior art OVE6D (w/o ICP) on LINEMOD, Occluded LINEMOD and T-LESS datasets respectively, and is comparable with other SOTA methods even the supervised ones. Compared with point-to-plane ICP, our method has the obvious advantage on computation speed, due to the merit of full play to
the high parallel characteristics of deep learning based on GPU acceleration.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)