Curvature Informed Furthest Point Sampling

Shubham Bhardwaj, Balakrisha Reddy Pailla

Keywords: dowsampling, point cloud, curvature informed, shape completion, segmentation, furthest point sampling

TL;DR: An extension of furthest point sampling algorithm that takes curvature information into consideration

Abstract: Point cloud representation is becoming increasingly popular due to its low memory footprint, ease of creation, collection and modification. As the size of the point cloud increases, we need to incorporate a down-sampling operation to meet the computational demands of the tasks. Classical approaches such as farthest point sampling perform exceedingly well over downstream tasks. The major drawback is that farthest point sampling is a mere heuristic and does not take geometric priors such as curvature into consideration. We propose a novel sampling procedure that conditions the output of farthest point sampling with curvature information. We create a joint rank by multiplying the soft furthest point rank with corresponding curvature scores obtained via a deep neural network and exchange a percentage of low-ranking points in the furthest set with the high-ranking points in the left-out set. Previous differentiable sampling approaches have failed to conform to the end-to-end learning paradigm due to instability while training. We demonstrate that our algorithm is compatible with end-to-end learning. Our sampling scheme consistently outperforms previous baselines on various downstream geometry processing tasks. Finally, we show detailed ablation studies regarding the qualitative and quantitative analysis of the role of different features used in the proposed algorithm.

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: Deep Learning and representational learning